Catastrophic Forgetting Mitigation via Discrepancy-Weighted Experience ReplaySupported by the National Key R &D Program of China under Grant 2022YFF0503900, and the Research Project of Institute of Software, Chinese Academy of Sciences (ISCAS-ZD-202401, ISCAS-ZD-202403).

Xinrun Xu, Jianwen Yang, Qiuhong Zhang, Zhanbiao Lian, Zhiming Ding, Shan Jiang

Abstract

Continually adapting edge models in cloud-edge collaborative object detection for traffic monitoring suffers from catastrophic forgetting, where models lose previously learned knowledge when adapting to new data distributions. This is especially problematic in dynamic traffic environments characterised by periodic variations (e.g., day/night, peak hours), where past knowledge remains valuable. Existing approaches like experience replay and visual prompts offer some mitigation, but struggle to effectively prioritize and leverage historical data for optimal knowledge retention and adaptation. Specifically, simply storing and replaying all historical data can be inefficient, while treating all historical experiences as equally important overlooks their varying relevance to the current domain. This paper proposes ER-EMU, an edge model update algorithm based on adaptive experience replay, to address these limitations. ER-EMU utilizes a limited-size experience buffer managed using a First-In-First-Out (FIFO) principle, and a novel Domain Distance Metric-based Experience Selection (DDM-ES) algorithm. DDM-ES employs the multi-kernel maximum mean discrepancy (MK-MMD) to quantify the dissimilarity between target domains, prioritizing the selection of historical data that is most dissimilar to the current target domain. This ensures training diversity and facilitates the retention of knowledge from a wider range of past experiences, while also preventing overfitting to the new domain. The experience buffer is also updated using a simple random sampling strategy to maintain a balanced representation of previous domains. Experiments on the Bellevue traffic video dataset, involving repeated day/night cycles, demonstrate that ER-EMU consistently improves the performance of several state-of-the-art cloud-edge collaborative object detection frameworks. Specifically, methods enhanced by ER-EMU show improved adaptation to repeated scenarios and significant performance gains over baselines, highlighting its effectiveness in mitigating catastrophic forgetting, improving adaptability, and enhancing model generalization in dynamic traffic monitoring scenarios. Furthermore, ablation studies show that DDM-ES plays a crucial role in the effectiveness of ER-EMU compared to a random historical experience selection, and the influence of the parameter $l$ on the ER-EMU algorithm is also experimentally analyzed. IoT \and Continual Learning \and Catastrophic Forgetting \and Edge Model Update \and Adaptive Experience Replay

Catastrophic Forgetting Mitigation via Discrepancy-Weighted Experience Replay††thanks: Supported by the National Key R&D Program of China under Grant 2022YFF0503900, and the Research Project of Institute of Software, Chinese Academy of Sciences (ISCAS-ZD-202401, ISCAS-ZD-202403).

Related Works

In cloud-edge collaborative frameworks, updating edge models to adapt to new environments is crucial but is plagued by catastrophic forgetting [14, 17]. As edge models continually adapt, they risk losing knowledge acquired from previous domains. This is especially problematic in traffic monitoring due to its cyclical patterns (e.g., day/night cycles), making historical knowledge valuable. The limitations of simply retraining from scratch with all data are resource constraints at the edge, possible overfitting, and impaired timeliness [18].

Existing work attempts to address these issues. The Shoggoth method [19] mitigates catastrophic forgetting through experience replay with a simple random sampling approach, and utilizes adaptive mini-batch training. Other methods, such as [13] propose gradient episodic memory to better preserve infor-

mation from past experiences. Aljundi et al. [1] approach this problem with a method that aims to learn what not to forget. However, these approaches do not distinguish the relative importance of different historical data points during the update. These methods treat all experiences equally, failing to identify when certain past data may be more pertinent for the current adaptation. DCC [6] addresses catastrophic forgetting using current data with unique visual prompts. While this has been shown to improve model generalization, it does not incorporate historical experiences for additional improvements.

This paper argues that strategically incorporating relevant historical information from different target domains is key to continual learning, which is particularly true in systems that display periodic changes in traffic patterns. The proposed ER-EMU method addresses these issues by adaptively selecting from historical data, ensuring a comprehensive learning experience while mitigating the risk of overfitting to new target domains.

Introduction

Continuously retraining edge models in cloud-edge-end collaborative architectures presents a significant challenge due to catastrophic forgetting [3-5, 10, 11, 15]. As these models adapt to new data distributions, they risk losing previously learned knowledge [14], potentially leading to inconsistent performance and difficulties in handling recurring scenarios [6, 20]. This is particularly critical in dynamic environments like traffic monitoring, where conditions such as day/night transitions and peak hour traffic patterns exhibit cyclical behavior, making previously learned knowledge highly valuable. The inability to retain this past knowledge impedes the model's ability to effectively adapt and generalize across these repeating but varying scenarios.

Retraining from scratch with all available data, while a direct solution [17,18], is often infeasible in resource-constrained edge environments. This approach also risks overfitting and may hinder the model's ability to generalize effectively, as the sheer volume of data could overwhelm the edge model's capacity for broad

learning. Furthermore, continuously retraining from scratch significantly undermines the timeliness of updates in these dynamic and time-sensitive environments. Recent efforts like Shoggoth [19] employ experience replay with adaptive mini-batch training to alleviate catastrophic forgetting. While this represents a step forward, such methods often lack a mechanism for discerning the relative importance of past experiences during retraining. Specifically, these methods often treat all historical data as equally important, failing to recognize that certain past experiences may be more relevant to the current adaptation task than others, and this could negatively impact the efficiency and effectiveness of the retraining process.

This paper proposes a novel edge model update algorithm based on adaptive experience replay, named ER-EMU, to address this limitation. Our algorithm leverages a tailored experience cache and management mechanism to selectively incorporate pertinent historical data during retraining. Unlike previous approaches, ER-EMU dynamically prioritizes past experiences using a domain distance metric, thus focusing on knowledge that is most relevant for current adaptation and thus preventing overfit to the current domain. This approach aims to mitigate catastrophic forgetting effectively, preserve historical knowledge, and improve the model's adaptability while remaining sensitive to the dynamic nature of edge environments [9,12,21]. The proposed method leverages both the recurring nature of traffic patterns and selectively leverages historic data to improve the performance of edge models.

We begin by formally defining the edge model update problem within the context of cloud-edge-end collaborations, focusing on the challenge of catastrophic forgetting in the context of continuous adaptation. Subsequently, we detail our proposed algorithm and its core components: the experience buffer, the Domain Distance Metric-based Experience Selection (DDM-ES) algorithm, and the random sampling based buffer update approach. Finally, we validate the efficacy of our method through extensive experiments conducted on real-world traffic video datasets, comparing its performance against state-of-the-art approaches and demonstrating the benefits of our adaptive replay mechanism.

Formal Description of the Edge Model Update Problem

This paper addresses the challenge of catastrophic forgetting in the context of a cloud-edge collaborative adaptive object detection framework. Our goal is to enable the lightweight edge model, denoted as θ edge , to maintain its performance on a continuously evolving sequence of target domains, D t 1 , D t 2 , ..., D tT , without sacrificing previously acquired knowledge. Each target domain, D ti = ( x i j , y i j ) j = 1 Nti , represents the data distribution encountered during the i -th model update cycle, corresponding to traffic monitoring video from a specific location and time period. Here, N ti represents the number of samples in D ti , while x i j and y i j denote the j -th sample and its corresponding label, respectively. The labels are generated by the cloud model, θ cloud .

The model learning paradigm involves leveraging knowledge acquired from the previous i -1 target domains to effectively adapt to the i -th target domain. Notably, the data distribution of each target domain can be arbitrary, although exhibiting some degree of periodicity in its evolution. The objective is formalized as shown in Equation 1,

where L is the loss function for the i -th target domain D ti , Φ is the regularization term to prevent catastrophic forgetting, and λ is the coefficient balancing their importance. The loss function L can be any suitable metric for object detection, such as mAP.

This study addresses the challenge of continual learning in cloud-edge collaborative object detection for traffic monitoring applications. Specifically, we aim to develop an effective edge model updating strategy that mitigates catastrophic forgetting while ensuring adaptability to evolving target domains and generalization across both historical and emerging data.

Edge Model Update Algorithm Based on Adaptive Experience Replay

ER-EMU, an edge model update algorithm that mitigates catastrophic forgetting during adaptation, leverages an experience replay mechanism (Figure 1). It maintains a FIFO experience buffer storing representative data from past target domains, used strategically during model fine-tuning. When adapting to a new target domain, the algorithm trains on a batch mixing data from the latest domain and the experience buffer. A dedicated sampling algorithm (Section 4.2) ensures the buffer's representativeness by selecting which new data points are incorporated.

Experience replay, a classic approach to alleviate catastrophic forgetting [1,13], leverages historical data or features during training. Its effectiveness depends on selecting relevant and diverse historical experiences (add citation). In our adaptive object detection scenario, we propose an experience selection algorithm based on inter-domain distance metrics. At each training round, the algorithm calculates the distance between the new target domain and all historical domains in the experience buffer. The m most distant historical domains, determined by this metric, augment the current target domain data during fine-tuning, promoting knowledge preservation from diverse domains. Among various distance metrics, Maximum Mean Discrepancy (MMD) [16] is widely used to quantify domain dissimilarity. Given domains D a and D b , represented by datasets X a = [ x a 1 , x a 2 , ..., X an a ] , and X b = [ x b 1 , x b 2 , ..., x bn b ] , their MMD is computed using Equation 2,

where ϕ ( · ) is a mapping function that maps the dataset X to a Hilbert space H (this step is usually achieved by using a specific kernel function k ( · ) ), and n a and n b represent the number of samples in the two domains, respectively.

However, MMD's reliance on a task-specific kernel function limits its generalizability. To address this, our approach utilizes Multi-Kernel Maximum Mean

Discrepancy (MK-MMD) [2,7] as the distance metric. MK-MMD combines multiple kernel functions (Equation 3) to represent the discrepancy between distributions, enhancing expressiveness and capturing more complex inter-domain relationships compared to traditional MMD. Consequently, MK-MMD has shown superior performance in cross-domain learning and pattern recognition tasks [2,7].

where K is the set of kernel functions, m is the number of kernel functions, β u is the weight of the u -th kernel function k u ( · ) , satisfying ∑ d u =1 β u = 1( β u ≥ 0) , and the specific form of K is shown in Equation 4.

Leveraging MK-MMD, we develop Domain Distance Metric-based Experience Selection (DDM-ES), detailed in Algorithm 1. If the experience buffer contains historical data, the algorithm computes the MK-MMD between each historical target domain and the current target domain, quantifying their dissimilarity. The l most distant historical domains form a set D H . Each domain in D H is then weighted using a Sigmoid function parameterized by their distances, ensuring that more dissimilar domains exert greater influence during training. This adaptive weighting in DDM-ES balances the incorporation of past experiences with mitigating overfitting to the current target domain.

Input: Experience buffer M , target domain dataset D ti . Output: Selected l historical target domain datasets D H = D h 1 , D h 2 , ..., D hl , and the weight of each historical target domain dataset W H = w h 1 , w h 2 , ..., w hl . 1: if IsEmpty(M) then 2: return 3: else 4: Initialize the output historical target domain set D H ← [ ] 5: Initialize the output weight set for historical target domains W H ← [ ] 6: Initialize the MK-MMD set dists between each historical target domain stored in M and D ti ← [ ] 7: for each historical target domain D hj stored in M do 8: Calculate the MK-MMD Dist MK ( D ti , D hj ) between D hj and D ti using Equation 3 9: Update dists , dists [ j ] ← Dist MK ( D ti , D hj ) 10: end for 11: Copy M , M copy ← M 12: Sort M copy and dists in ascending order based on dists , ( dists , M copy ) ← sort ( M copy , dists ) 13: Update D H ← M copy [: l ] 14: for each historical target domain D j in D H do 15: Update the corresponding weight W H [ j ] ← sigmoid ( dists [ j ]) 16: end for 17: end if

Algorithm 1 Domain Distance Metric-based Experience Selection (DDM-ES)

To maximize data diversity, our approach enforces balanced representation within the experience buffer, storing an equal number of samples from each target domain. Assuming independent and identically distributed data within each domain, a random sampling strategy selects data points for inclusion. This facilitates equitable knowledge sharing across domains (Algorithm 2). When processing a new target domain, if the buffer is full, random samples are removed to create space. Then, random samples from the new domain are incorporated, ensuring the buffer's capacity limit and maximizing knowledge retention from all encountered domains.

Input: Experience buffer M , the number of images h to be saved per domain, target domain dataset D ti Output: None 1: if IsFull( M ) then 2: Randomly sample h images M add from D ti 3: Select the oldest h images M replace from M 4: Update M ← ( M -M replace ) ∪ M add 5: else 6: Randomly sample h images M add from D ti 7: Update M ← M ∪ M add 8: end if

Integrating the DDM-ES and RS-EBU algorithms, we propose the Edge Model Update Algorithm Based on Adaptive Experience Replay (ER-EMU). Upon arrival of a new target domain, D ti , DDM-ES identifies a subset of l most dissimilar historical target domain datasets, D H = D h 1 , D h 2 , ..., D hl , and computes their corresponding weight set W H = w h 1 , w h 2 , ..., w hl , reflecting their training importance. The loss function for this round of edge model updates is then formulated as shown in Equation 5.

After the edge model updates using the combined data, RS-EBU incorporates the new target domain, D ti , into the experience buffer, ensuring its representativeness. ER-EMU optimizes the objective function (Equation 1) by strategically recalling relevant past knowledge during each model adaptation cycle, mitigating catastrophic forgetting and enabling the edge model to maintain performance on previously learned domains while adapting to new ones.

The experimental evaluation validates the effectiveness of ER-EMU within a cloud-edge collaborative architecture. Our methodology integrates ER-EMU into several SOTA cloud-edge collaborative object detection frameworks, replacing their native edge model update modules. By comparing performance on realworld traffic monitoring data before and after incorporating ER-EMU, we assess its efficacy in mitigating catastrophic forgetting and enhancing adaptability.

Experimental Setup

Our experimental setup uses a pre-trained DINO [22] model (ResNet-101 [8] backbone) for the cloud server and a lighter RT-DETR model [23] (ResNet-18 backbone) for the edge server. The Bellevue traffic video dataset provides realworld traffic scenarios for evaluation. Table 1 summarizes the hyperparameters used in our experiments.

Table 1: Experimental hyperparameters for performance validation of ER-EMU algorithm.

Baselines: 1) Shoggoth [19]: Cloud-edge, random historical data sampling, combats catastrophic forgetting. 2) Shoggoth-ER-EMU: Shoggoth enhanced with ER-EMU edge model updates. 3) DCC [6]: Cloud-edge, current domain data, visual prompts for forgetting mitigation. 4) DCC-ER-EMU: DCC enhanced with ER-EMU edge model updates. 5) LVACCL: Cloud-edge, video stream correlations, edge model retraining. 6) LVACCL-ER-EMU: LVACCL with ER-EMU, modified temporal correlation handling.

Effectiveness of the ER-EMU Algorithm in Mitigating Catastrophic Forgetting in Edge Models

To validate ER-EMU's effectiveness in addressing catastrophic forgetting, we tested it in an extreme scenario: cycling through two daytime/nighttime Bellevue traffic video scenes twice. ER-EMU was applied to Shoggoth, DCC, and LVACCL. Table 2 shows average mAP for each method per scene and overall. Results indicate that ER-EMU consistently improves detection accuracy for SOTA cloud-edge adaptive methods. Methods using ER-EMU show improved performance upon re-encountering scenes, demonstrating its effectiveness in mitigating

Table 2: Continual Adaptation capability on Bellevue Traffic with repeat scene for ER-EMU.

catastrophic forgetting. While Shoggoth and DCC (mini-batch experience replay & visual prompts) show some resistance, ER-EMU further improves their performance, indicating its superiority over the comparison methods' edge model update modules in mitigating catastrophic forgetting.

Performance Validation Experiments for the ER-EMU Algorithm

• (a) Experimental results of ablation study for ER-EMU.
• (b) Parameter l on the performance of ER-EMU.

Ablation Study

To validate ER-EMU's edge model update performance, we applied the methods to 8 Bellevue camera video streams with varying time spans. Results in Figure 2 (1, 2, 4, 7 day spans) demonstrate ER-EMU enhances SOTA cloud-edge methods and exhibits robustness across time spans. This highlights the importance of leveraging past experiences and dynamically updating environmental models.

ER-EMU's smaller improvement on DCC is likely due to DCC's specific catastrophic forgetting optimizations and visual prompts. However, DCC still improved with ER-EMU, indicating ER-EMU's superiority.

Figure 4 shows detection accuracy gains across time spans. Figures 4 (Shoggoth, DCC, LVACCL) depict gain changes. While gains vary across methods, the pattern of increasing gain with longer time spans is similar, suggesting ER-EMU is more advantageous for longer streams. This relates to its ability to cache key historical samples and adaptively select data for retraining. ER-EMU achieves continuous gains in traffic environments with regular variations. However, due to limited model capacity, gains eventually slow down.

Acore module of the ER-EMU algorithm lies in using the DDM-ES algorithm to select samples from the experience buffer for retraining the edge model in a domain distance-based manner. This section focuses on investigating the role of this algorithm within the overall ER-EMU algorithm.

To isolate ER-EMU's components, a trimmed ER-EMU (DDM-ES replaced with random sampling) and the full ER-EMU were applied to Shoggoth, DCC, and LVACCL using a week of video data from 8 Bellevue cameras. Figure 4a shows that the random sampling ER-EMU still improves performance, suggesting random historical data selection can alleviate catastrophic forgetting. The performance decrease of random sampling ER-EMU compared to the full EREMU is most significant for Shoggoth. This is likely because Shoggoth already uses random sampling for experience buffer updates, only lacking a priority-based selection mechanism; its update algorithm is similar to the trimmed ER-EMU. This, combined with the other methods' results, demonstrates the importance of the DDM-ES historical experience selection algorithm in ER-EMU.

To investigate the influence of the hyperparameter l in the ER-EMU algorithm, experiments were conducted with l values ranging from 1 to 10, keeping all other parameters consistent. Results, as depicted in Figure 4b, demonstrate that variations in l do not significantly impact the algorithm's performance. Therefore, the ER-EMU algorithm exhibits robustness and low sensitivity to the choice of the hyperparameter l .

Conclusion

This paper addresses the challenge of catastrophic forgetting in edge model retraining for cloud-edge collaborative object detection in traffic monitoring. It proposes ER-EMU, an edge model update algorithm based on adaptive experience replay. ER-EMU utilizes a limited-size experience buffer and a novel experience selection algorithm based on inter-domain distance metrics. This approach selects the most representative historical data, enriching the training set, improving generalization, and mitigating forgetting. The experience selection leverages multi-kernel maximum mean discrepancy to quantify domain distance, prioritizing diverse historical data for replay. Additionally, a random sampling-based buffer update mechanism maintains data diversity and balance. Experiments on the Bellevue traffic video dataset demonstrate ER-EMU's superiority over existing edge model update algorithms in various cloud-edge collaborative architectures.

Algorithm: Domain Distance Metric-based Experience Selection (DDM-ES)

\scriptsize

\Statex \hspace*-\algorithmicindent Input: Experience buffer $M$, target domain dataset $D_ti$.
\Statex \hspace*-\algorithmicindent Output: Selected $l$ historical target domain datasets $D_H = \{D_h1, D_h2, ... , D_hl\}$, and the weight of each historical target domain dataset $W_H = \{w_h1, w_h2, ... , w_hl\}$.
 **if** IsEmpty(M) **then**
 **return** **else**
 Initialize the output historical target domain set $D_H \gets [\ ]$
 
 Initialize the output weight set for historical target domains $W_H \gets [\ ]$
 
 Initialize the MK-MMD set $dists$ between each historical target domain stored in $M$ and $D_ti \gets [\ ]$
 
 **for** each historical target domain $D_hj **do**$ stored in $M$
 Calculate the MK-MMD $Dist_MK(D_ti, D_hj)$ between $D_hj$ and $D_ti$ using Equation 
 Update $dists$, $dists[j] \gets Dist_MK(D_ti, D_hj)$
 **end for** 
 Copy $M$, $M_copy \gets M$ 
 Sort $M_copy$ and $dists$ in ascending order based on $dists$, ($dists$,$M_copy$) $\gets sort(M_copy, dists)$
 Update $D_H \gets M_copy[:l]$
 **for** each historical target domain $D_j **do**$ in $D_H$
 Update the corresponding weight $W_H[j] \gets sigmoid(dists[j])$
 **end for**
 **end if**

Algorithm: Random Sampling-based Experience Buffer Update Algorithm (RS-EBU)

\scriptsize

\Statex \hspace*-\algorithmicindent Input: Experience buffer $M$, the number of images $h$ to be saved per domain, target domain dataset $D_ti$
\Statex \hspace*-\algorithmicindent Output: None
 **if** IsFull($M$) **then** 
 Randomly sample $h$ images $M_\textadd$ from $D_ti$
 Select the oldest $h$ images $M_\textreplace$ from $M$
 Update $M \gets (M - M_\textreplace) \cup M_\textadd$
 **else**
 Randomly sample $h$ images $M_\textadd$ from $D_ti$ 
 Update $M \gets M \cup M_\textadd$ 
 **end if**

References

Continually adapting edge models in cloud-edge collaborative object detection for traffic monitoring suffers from catastrophic forgetting, where models lose previously learned knowledge when adapting to new data distributions. This is especially problematic in dynamic traffic environments characterised by periodic variations (e.g., day/night, peak hours), where past knowledge remains valuable. Existing approaches like experience replay and visual prompts offer some mitigation, but struggle to effectively prioritize and leverage historical data for optimal knowledge retention and adaptation. Specifically, simply storing and replaying all historical data can be inefficient, while treating all historical experiences as equally important overlooks their varying relevance to the current domain. This paper proposes ER-EMU, an edge model update algorithm based on adaptive experience replay, to address these limitations. ER-EMU utilizes a limited-size experience buffer managed using a First-In-First-Out (FIFO) principle, and a novel Domain Distance Metric-based Experience Selection (DDM-ES) algorithm. DDM-ES employs the multi-kernel maximum mean discrepancy (MK-MMD) to quantify the dissimilarity between target domains, prioritizing the selection of historical data that is most dissimilar to the current target domain. This ensures training diversity and facilitates the retention of knowledge from a wider range of past experiences, while also preventing overfitting to the new domain. The experience buffer is also updated using a simple random sampling strategy to maintain a balanced representation of previous domains. Experiments on the Bellevue traffic video dataset, involving repeated day/night cycles, demonstrate that ER-EMU consistently improves the performance of several state-of-the-art cloud-edge collaborative object detection frameworks. Specifically, methods enhanced by ER-EMU show improved adaptation to repeated scenarios and significant performance gains over baselines, highlighting its effectiveness in mitigating catastrophic forgetting, improving adaptability, and enhancing model generalization in dynamic traffic monitoring scenarios. Furthermore, ablation studies show that DDM-ES plays a crucial role in the effectiveness of ER-EMU compared to a random historical experience selection, and the influence of the parameter l on the ER-EMU algorithm is also experimentally analyzed.

Retraining from scratch with all available data, while a direct solution [17, 18], is often infeasible in resource-constrained edge environments. This approach also risks overfitting and may hinder the model’s ability to generalize effectively, as the sheer volume of data could overwhelm the edge model’s capacity for broad learning. Furthermore, continuously retraining from scratch significantly undermines the timeliness of updates in these dynamic and time-sensitive environments. Recent efforts like Shoggoth [19] employ experience replay with adaptive mini-batch training to alleviate catastrophic forgetting. While this represents a step forward, such methods often lack a mechanism for discerning the relative importance of past experiences during retraining. Specifically, these methods often treat all historical data as equally important, failing to recognize that certain past experiences may be more relevant to the current adaptation task than others, and this could negatively impact the efficiency and effectiveness of the retraining process.

Existing work attempts to address these issues. The Shoggoth method [19] mitigates catastrophic forgetting through experience replay with a simple random sampling approach, and utilizes adaptive mini-batch training. Other methods, such as [13] propose gradient episodic memory to better preserve information from past experiences. Aljundi et al. [1] approach this problem with a method that aims to learn what not to forget. However, these approaches do not distinguish the relative importance of different historical data points during the update. These methods treat all experiences equally, failing to identify when certain past data may be more pertinent for the current adaptation. DCC [6] addresses catastrophic forgetting using current data with unique visual prompts. While this has been shown to improve model generalization, it does not incorporate historical experiences for additional improvements.

Table: S5.T1: Experimental hyperparameters for performance validation of ER-EMU algorithm.

Hyperparameter	Setting	Description
Ms​i​z​e	30	Maximum buffer size for historical target domains.
h	200	Samples per historical domain.
l	5	Historical domains selected per update.

Table: S5.T2: Continual Adaptation capability on Bellevue Traffic with repeat scene for ER-EMU.

Video	day1	day2	night1	night2	day1	day2	night1	night2	Mean
Shoggoth	61.9	59.7	57.5	58.6	62.4	60.3	58.9	58.9	59.8
Shoggoth-ER-EMU	63.1	61.3	59.7	60.2	64.7	63.8	62.8	62.2	62.2
DCC	60.7	60.5	56.7	57.3	62.7	59.4	57.7	55.4	58.9
DCC-ER-EMU	63.9	63.5	62.3	63.1	65.5	65.3	62.5	63.1	63.7
LVCCL	62.5	59.4	56.3	57.2	61.5	61.3	56.9	58.7	59.2
LVCCL-ER-EMU	63.6	60.0	57.0	58.2	64.2	61.7	58.2	59.1	60.3

$$ \displaystyle H_{MMD}= $$

$$ \displaystyle+\frac{1}{n_{b}^{2}}\sum_{j=1}^{n_{b}}\sum_{j^{\prime}=1}^{n_{b}}\langle\phi(x_{bj}),\phi(x_{bj^{\prime}})\rangle_{\mathcal{H}} $$

Time	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →	t - -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --- →
Video	day1	day2	night1	night2	day1	day2	night1	night2	Mean
Shoggoth	61.9	59.7	57.5	58.6	62.4	60.3	58.9	58.9	59.8
Shoggoth-ER-EMU	63.1	61.3	59.7	60.2	64.7	63.8	62.8	62.2	62.2
DCC	60.7	60.5	56.7	57.3	62.7	59.4	57.7	55.4	58.9
DCC-ER-EMU	63.9	63.5	62.3	63.1	65.5	65.3	62.5	63.1	63.7
LVCCL	62.5	59.4	56.3	57.2	61.5	61.3	56.9	58.7	59.2
LVCCL-ER-EMU	63.6	60.0	57.0	58.2	64.2	61.7	58.2	59.1	60.3

References

[1] Ganin, Yaroslav, Ustinova, Evgeniya, Ajakan, Hana, Germain, Pascal, Larochelle, Hugo, Laviolette, Fran{\c{c. (2016). Domain-adversarial training of neural networks. Journal of machine learning research.

[2] Oza, Poojan, Sindagi, Vishwanath A, Sharmini, Vibashan Vishnukumar, Patel, Vishal M. (2023). Unsupervised domain adaptation of object detectors: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence.

[3] Smith, James Seale, Karlinsky, Leonid, Gutta, Vyshnavi, Cascante-Bonilla, Paola, Kim, Donghyun, Arbelle, Assaf, Panda, Rameswar, Feris, Rogerio, Kira, Zsolt. (2023). Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

[4] Woo, Sanghyun, Park, Jongchan, Lee, Joon-Young, Kweon, In So. (2018). Cbam: Convolutional block attention module. Proceedings of the European conference on computer vision (ECCV).

[5] Gan, Yulu, Bai, Yan, Lou, Yihang, Ma, Xianzheng, Zhang, Renrui, Shi, Nian, Luo, Lin. (2023). Decorate the newcomers: Visual domain prompt for continual test time adaptation. Proceedings of the AAAI Conference on Artificial Intelligence.

[6] Jia, Menglin, Tang, Luming, Chen, Bor-Chun, Cardie, Claire, Belongie, Serge, Hariharan, Bharath, Lim, Ser-Nam. (2022). Visual prompt tuning. European Conference on Computer Vision.

[7] Xing, Wan, Sultan Mohd, Mohd Rizman, Johari, Juliana, Ahmat Ruslan, Fazlina. (2023). A review on object detection algorithms based deep learning methods. Journal of Electrical and Electronic Systems Research (JEESR).

[8] Liu, Shilong, Zeng, Zhaoyang, Ren, Tianhe, Li, Feng, Zhang, Hao, Yang, Jie, Li, Chunyuan, Yang, Jianwei, Su, Hang, Zhu, Jun, others. (2023). Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499.

[9] Krichen, Moez. (2023). Deep reinforcement learning. 2023 14th International Conference on Computing Communication and Networking Technologies (ICCCNT).

[10] 谈海生，张欣，郑子木，李向阳. (2023). 边缘计算理论与系统实践：基于CNCF KubeEdge的实现.

[11] Wang, Wen, Cao, Yang, Zhang, Jing, He, Fengxiang, Zha, Zheng-Jun, Wen, Yonggang, Tao, Dacheng. (2021). Exploring sequence feature alignment for domain adaptive detection transformers. Proceedings of the 29th ACM International Conference on Multimedia.

[12] Wu, Jinglong, Xue, Chao, Zhao, Ziran, Liu, Bohang. (2012). Effect of camera angle on precision of parameters for traffic accident in video detection. 2012 Third International Conference on Digital Manufacturing & Automation.

[13] Ho, Stella, Liu, Ming, Du, Lan, Gao, Longxiang, Xiang, Yong. (2023). Prototype-guided memory replay for continual learning. IEEE Transactions on Neural Networks and Learning Systems.

[14] Kong, Yajing, Liu, Liu, Chen, Huanhuan, Kacprzyk, Janusz, Tao, Dacheng. (2023). Overcoming catastrophic forgetting in continual learning by exploring eigenvalues of Hessian matrix. IEEE Transactions on Neural Networks and Learning Systems.

[15] Lamers, Christiaan, Vidal, Ren{'e. (2023). Clustering-based domain-incremental learning. Proceedings of the IEEE/CVF International Conference on Computer Vision.

[16] Wang, Jiaqi, Liu, Zhengliang, Zhao, Lin, Wu, Zihao, Ma, Chong, Yu, Sigang, Dai, Haixing, Yang, Qiushi, Liu, Yiheng, Zhang, Songyao, others. (2023). Review of large vision models and visual prompt engineering. Meta-Radiology.

[17] Li, Jing, Guo, Song, Liang, Weifa, Wang, Jianping, Chen, Quan, Zeng, Yue, Ye, Baoliu, Jia, Xiaohua. (2023). Digital Twin-Enabled Service Provisioning in Edge Computing via Continual Learning. IEEE Transactions on Mobile Computing.

[18] Shao, Yunli, Wang, Chieh, Berres, Andy, Yoshioka, Jovan, Cook, Adian, Xu, Haowen. (2022). Computer vision-enabled smart traffic monitoring for sustainable transportation management. International Conference on Transportation and Development 2022.

[19] Zhou, Xuan, Ke, Ruimin, Yang, Hao, Liu, Chenxi. (2021). When intelligent transportation systems sensing meets edge computing: Vision and challenges. Applied Sciences.

[20] Li, Dawei, Tasci, Serafettin, Ghosh, Shalini, Zhu, Jingwen, Zhang, Junting, Heck, Larry. (2019). RILOD: Near real-time incremental learning for object detection at the edge. Proceedings of the 4th ACM/IEEE Symposium on Edge Computing.

[21] Ding, Chuntao, Zhou, Ao, Liu, Yunxin, Chang, Rong N, Hsu, Ching-Hsien, Wang, Shangguang. (2020). A cloud-edge collaboration framework for cognitive service. IEEE Transactions on Cloud Computing.

[22] Davalas, Charalampos, Michail, Dimitrios, Varlamis, Iraklis, Tserpes, Konstantinos. (2022). A Cloud-based Continual Learning System for Road Sign Classification in Autonomous Driving.. CI4PM/PAI@ WCCI.

[23] Liu, Guanxiong, Shi, Hang, Kiani, Abbas, Khreishah, Abdallah, Lee, Joyoung, Ansari, Nirwan, Liu, Chengjun, Yousef, Mustafa Mohammad. (2021). Smart traffic monitoring system using computer vision and edge computing. IEEE Transactions on Intelligent Transportation Systems.

[24] Skadins, Ansis, Ivanovs, Maksims, Rava, Raimonds, Nesenbergs, Krisjanis. (2020). Edge pre-processing of traffic surveillance video for bandwidth and privacy optimization in smart cities. 2020 17th Biennial Baltic Electronics Conference (BEC).

[25] Al-qaness, Mohammed AA, Abbasi, Aaqif Afzaal, Fan, Hong, Ibrahim, Rehab Ali, Alsamhi, Saeed H, Hawbani, Ammar. (2021). An improved YOLO-based road traffic monitoring system. Computing.

[26] Zhang, Bo, Zhang, Jian. (2020). A traffic surveillance system for obtaining comprehensive information of the passing vehicles based on instance segmentation. IEEE Transactions on Intelligent Transportation Systems.

[27] Zhang, Zhimei, Liu, Kun, Gao, Feng, Li, Xianyun, Wang, Guodong. (2016). Vision-based vehicle detecting and counting for traffic flow analysis. 2016 International Joint Conference on Neural Networks (IJCNN).

[28] Ghahremannezhad, Hadi, Shi, Hang, Liu, Chengjun. (2023). Object detection in traffic videos: A survey. IEEE Transactions on Intelligent Transportation Systems.

[29] Han, Yajing, Hu, Dean. (2020). Multispectral fusion approach for traffic target detection in bad weather. Algorithms.

[30] Shi, Yifeng, Lv, Feng, Wang, Xinliang, Xia, Chunlong, Li, Shaojie, Yang, Shujie, Xi, Teng, Zhang, Gang. (2023). Open-transmind: A new baseline and benchmark for 1st foundation model challenge of intelligent transportation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

[31] Guo, Xiangyu, Gao, Mingliang, Zhai, Wenzhe, Li, Qilei, Jeon, Gwanggil. (2023). Scale region recognition network for object counting in intelligent transportation system. IEEE Transactions on Intelligent Transportation Systems.

[32] Sharma, Neerav, Garg, Rahul D. (2023). Real-time iot-based connected vehicle infrastructure for intelligent transportation safety. IEEE Transactions on Intelligent Transportation Systems.

[33] Redmon, Joseph, Divvala, Santosh, Girshick, Ross, Farhadi, Ali. (2016). You only look once: Unified, real-time object detection. Proceedings of the IEEE conference on computer vision and pattern recognition.

[34] Ren, Shaoqing, He, Kaiming, Girshick, Ross, Sun, Jian. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems.

[35] Khani, Mehrdad, Hamadanian, Pouya, Nasr-Esfahany, Arash, Alizadeh, Mohammad. (2021). Real-time video inference on edge devices via adaptive model streaming. Proceedings of the IEEE/CVF International Conference on Computer Vision.

[36] Wang, Liang, Lu, Kai, Zhang, Nan, Qu, Xiaoyang, Wang, Jianzong, Wan, Jiguang, Li, Guokuan, Xiao, Jing. (2023). Shoggoth: towards efficient edge-cloud collaborative real-time video inference via adaptive online learning. 2023 60th ACM/IEEE Design Automation Conference (DAC).

[37] Gan, Yulu, Pan, Mingjie, Zhang, Rongyu, Ling, Zijian, Zhao, Lingran, Liu, Jiaming, Zhang, Shanghang. (2023). Cloud-device collaborative adaptation to continual changing environments in the real-world. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

[38] Nan, Ya, Jiang, Shiqi, Li, Mo. (2023). Large-scale Video Analytics with Cloud--Edge Collaborative Continuous Learning. ACM Transactions on Sensor Networks.

[39] Zhang, Miao, Wang, Fangxin, Liu, Jiangchuan. (2022). Casva: Configuration-adaptive streaming for live video analytics. IEEE INFOCOM 2022-IEEE Conference on Computer Communications.

[40] Bhardwaj, Romil, Xia, Zhengxu, Ananthanarayanan, Ganesh, Jiang, Junchen, Shu, Yuanchao, Karianakis, Nikolaos, Hsieh, Kevin, Bahl, Paramvir, Stoica, Ion. (2022). Ekya: Continuous learning of video analytics models on edge compute servers. 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22).

[41] Wang, Liang, Zhang, Nan, Qu, Xiaoyang, Wang, Jianzong, Wan, Jiguang, Li, Guokuan, Hu, Kaiyu, Jiang, Guilin, Xiao, Jing. (2023). EdgeMA: Model Adaptation System for Real-Time Video Analytics on Edge Devices. International Conference on Neural Information Processing.

[42] Chen, Bo, Bakhshi, Ali, Batista, Gustavo, Ng, Brian, Chin, Tat-Jun. (2022). Update compression for deep neural networks on the edge. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

[43] Dai, Cheng, Liu, Xingang, Chen, Weiting, Lai, Chin-Feng. (2020). A low-latency object detection algorithm for the edge devices of IoV systems. IEEE Transactions on Vehicular Technology.

[44] Vaswani, Ashish, Shazeer, Noam, Parmar, Niki, Uszkoreit, Jakob, Jones, Llion, Gomez, Aidan N, Kaiser, {\L. (2017). Attention is all you need. Advances in neural information processing systems.

[45] Carion, Nicolas, Massa, Francisco, Synnaeve, Gabriel, Usunier, Nicolas, Kirillov, Alexander, Zagoruyko, Sergey. (2020). End-to-end object detection with transformers. European conference on computer vision.

[46] Ying, Xue. (2019). An overview of overfitting and its solutions. Journal of physics: Conference series.

[47] McCloskey, Michael, Cohen, Neal J. (1989). Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of learning and motivation.

[48] Sutton, Richard S, Barto, Andrew G. (1999). Reinforcement learning: An introduction. Robotica.

[49] Monahan, George E. (1982). State of the art—a survey of partially observable Markov decision processes: theory, models, and algorithms. Management science.

[50] Littman, Michael L. (1994). Markov games as a framework for multi-agent reinforcement learning. Machine learning proceedings 1994.

[51] Shapley, Lloyd S. (1953). Stochastic games. Proceedings of the national academy of sciences.

[52] Lv, Wenyu, Xu, Shangliang, Zhao, Yian, Wang, Guanzhong, Wei, Jinman, Cui, Cheng, Du, Yuning, Dang, Qingqing, Liu, Yi. (2023). Detrs beat yolos on real-time object detection. arXiv preprint arXiv:2304.08069.

[53] Caron, Mathilde, Touvron, Hugo, Misra, Ishan, J{'e. (2021). Emerging properties in self-supervised vision transformers. Proceedings of the IEEE/CVF international conference on computer vision.

[54] Chen, Kai, Wang, Jiaqi, Pang, Jiangmiao, Cao, Yuhang, Xiong, Yu, Li, Xiaoxiao, Sun, Shuyang, Feng, Wansen, Liu, Ziwei, Xu, Jiarui, others. (2019). MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155.

[55] P. Sadhukhan, Sahali Banerjee, P. Das. (2021). Road Traffic Congestion Monitoring in Urban Areas: A Review. doi:10.1007/978-3-030-70183-3_8.

[56] Dimitrakopoulos, George, Demestichas, Panagiotis. (2010). Intelligent transportation systems. IEEE Vehicular Technology Magazine.

[57] Wang, Fei-Yue, Lin, Yilun, Ioannou, Petros A, Vlacic, Ljubo, Liu, Xiaoming, Eskandarian, Azim, Lv, Yisheng, Na, Xiaoxiang, Cebon, David, Ma, Jiaqi, others. (2023). Transportation 5.0: The DAO to safe, secure, and sustainable intelligent transportation systems. IEEE Transactions on Intelligent Transportation Systems.

[58] Du, Fu-Jun, Jiao, Shuang-Jian. (2022). Improvement of lightweight convolutional neural network model based on YOLO algorithm and its research in pavement defect detection. Sensors.

[59] Madhuri, T, Sowjanya, P. (2016). Microsoft Azure v/s Amazon AWS cloud services: A comparative study. International Journal of Innovative Research in Science, Engineering and Technology.

[60] Ali, Muhammad, Anjum, Ashiq, Rana, Omer, Zamani, Ali Reza, Balouek-Thomert, Daniel, Parashar, Manish. (2020). RES: Real-time video stream analytics using edge enhanced clouds. IEEE Transactions on Cloud Computing.

[61] Satyanarayanan, Mahadev. (2017). The emergence of edge computing. Computer.

[62] Shubha, Sudipta Saha, Shen, Haiying. (2023). AdaInf: Data Drift Adaptive Scheduling for Accurate and SLO-guaranteed Multiple-Model Inference Serving at Edge Servers. Proceedings of the ACM SIGCOMM 2023 Conference.

[63] Gu, Zhenghao, Lei, Yanjing, Chan, Sixian, Cao, Di, Zhang, Kongkai. (2022). Traffic Incident Detection System Based on Video Analysis. Proceedings of the 2022 5th International Conference on Computational Intelligence and Intelligent Systems.

[64] Li, Min, Li, Yu, Tian, Ye, Jiang, Li, Xu, Qiang. (2021). AppealNet: An efficient and highly-accurate edge/cloud collaborative architecture for DNN inference. 2021 58th ACM/IEEE Design Automation Conference (DAC).

[65] Liu, Ruoyang, Zhang, Lu, Wang, Jingyu, Yang, Huazhong, Liu, Yongpan. (2021). PETRI: Reducing bandwidth requirement in smart surveillance by edge-cloud collaborative adaptive frame clustering and pipelined bidirectional tracking. 2021 58th ACM/IEEE Design Automation Conference (DAC).

[66] Liu, Weihong, Geng, Jiawei, Zhu, Zongwei, Cao, Jing, Lian, Zirui. (2022). Sniper: Cloud-edge collaborative inference scheduling with neural network similarity modeling. Proceedings of the 59th ACM/IEEE Design Automation Conference.

[67] Chen, Chen, Liu, Bin, Wan, Shaohua, Qiao, Peng, Pei, Qingqi. (2020). An edge traffic flow detection scheme based on deep learning in an intelligent transportation system. IEEE Transactions on Intelligent Transportation Systems.

[68] Wu, Libing, Zhang, Rui, Zhou, Ruiting, Wu, Dan. (2021). An edge computing based data detection scheme for traffic light at intersections. Computer Communications.

[69] Khan, Asif, Khattak, Khurram S, Khan, Zawar H, Gulliver, Thomas Aaron, Abdullah. (2023). Edge computing for effective and efficient traffic characterization. Sensors.

[70] Wu, Duo, Zhang, Dayou, Zhang, Miao, Zhang, Ruoyu, Wang, Fangxin, Cui, Shuguang. (2023). Ilcas: Imitation learning-based configuration-adaptive streaming for live video analytics with cross-camera collaboration. IEEE Transactions on Mobile Computing.

[71] Ryu, Heechang, Shin, Hayong, Park, Jinkyoo. (2020). Multi-agent actor-critic with hierarchical graph attention network. Proceedings of the AAAI Conference on Artificial Intelligence.

[72] Hu, Jie, Shen, Li, Sun, Gang. (2018). Squeeze-and-excitation networks. Proceedings of the IEEE conference on computer vision and pattern recognition.

[73] Paszke, Adam, Gross, Sam, Massa, Francisco, Lerer, Adam, Bradbury, James, Chanan, Gregory, Killeen, Trevor, Lin, Zeming, Gimelshein, Natalia, Antiga, Luca, others. (2019). Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems.

[74] Abadi, Mart{'\i. (2016). *${$TensorFlow$*. 12th USENIX symposium on operating systems design and implementation (OSDI 16).

[75] Wang, Chengjia, Yang, Guang, Papanastasiou, Giorgos, Zhang, Heye, Rodrigues, Joel JPC, De Albuquerque, Victor Hugo C. (2020). Industrial cyber-physical systems-based cloud IoT edge for federated heterogeneous distillation. IEEE Transactions on Industrial Informatics.

[76] Wikibook. (2014). http://en.wikibooks.org/wiki/LaTeX.

[77] Lamport, Leslie. (1986). Document Preparation System.

[78] 陈浩元. (2005). 著录文后参考文献的规则及注意事项. 编辑学报.

[79] 初景利. (2004). 图书馆数字参考咨询服务研究.

[80] Stamerjohanns, Heinrich, Ginev, Deyan, David, Catalin, Misev, Dimitar, Zamdzhiev, Vladimir, Kohlhase, Michael. (2009). {MathML. Towards a Digital Mathematics Library.

[81] Betts, Lisa R, Taylor, Christopher P. (2005). Aging reduces center-surround antagonism in visual motion processing. Neuron.

[82] Bravo, Hermes, Olavarria, Jaime. (1990). Comparative study of visual inter and intrahemispheric cortico-cortical connections in five native Chilean rodents. Anatomy and embryology.

[83] 哈里森·沃尔德伦. (2012). 经济数学与金融数学.

[84] . 综合湿地管理国际研讨会论文集. (2013).

[85] 陈晋镳, 张惠民, 朱士兴, 赵震, 王振刚. (1980). 蓟县震旦亚界研究. 中国震旦亚界.

[86] 袁训来, 陈哲, 肖书海. (2012). 蓝田生物群: 一个认识多细胞生物起源和早期演化的新窗口 -- 篇一. 科学通报.

[87] 袁训来, 陈哲, 肖书海. (2012). 蓝田生物群: 一个认识多细胞生物起源和早期演化的新窗口 -- 篇二. 科学通报.

[88] 袁训来, 陈哲, 肖书海. (2012). 蓝田生物群: 一个认识多细胞生物起源和早期演化的新窗口 -- 篇三. 科学通报.

[89] Walls, Susan C., Barichivich, William J., Brown, Mary E.. (2013). Drought, deluge and declines: the impact of precipitation extremes on amphibians in a changing climate. Biology. doi:10.3390/biology2010399.

[90] ボハン, デ. (1928). 過去及び現在に於ける英国と会. 日本時報.

[91] Дубровин, А. И. (1906). Открытое письмо Председателя Главного Совета Союза Русского Народа Санкт-Петербургскому Антонию, Первенствующему члену Священного Синода. Вече.

[92] Parisi, German I, Kemker, Ronald, Part, Jose L, Kanan, Christopher, Wermter, Stefan. (2019). Continual lifelong learning with neural networks: A review. Neural networks.

[93] Rebuffi, Sylvestre-Alvise, Kolesnikov, Alexander, Sperl, Georg, Lampert, Christoph H. (2017). icarl: Incremental classifier and representation learning. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition.

[94] Lopez-Paz, David, Ranzato, Marc'Aurelio. (2017). Gradient episodic memory for continual learning. Advances in neural information processing systems.

[95] Aljundi, Rahaf, Babiloni, Francesca, Elhoseiny, Mohamed, Rohrbach, Marcus, Tuytelaars, Tinne. (2018). Memory aware synapses: Learning what (not) to forget. Proceedings of the European conference on computer vision (ECCV).

[96] Gretton, Arthur, Borgwardt, Karsten M, Rasch, Malte J, Sch{. (2012). A kernel two-sample test. The Journal of Machine Learning Research.

[97] Cortes, Corinna, Mohri, Mehryar, Rostamizadeh, Afshin. (2012). Algorithms for learning kernels based on centered alignment. The Journal of Machine Learning Research.

[98] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, Heung-Yeung Shum. (2022). DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection.

[99] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. (2015). Deep Residual Learning for Image Recognition.

[100] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, Jie Chen. (2024). DETRs Beat YOLOs on Real-time Object Detection.

[101] Ding, Zhiming, Jiang, Shan, Xu, Xinrun, Han, Yanbo. (2022). An Internet of Things based scalable framework for disaster data management. Journal of safety science and resilience.

[102] Ding, Zhiming, Xu, Xinrun, Jiang, Shan, Yan, Jin, Han, Yanbo. (2022). Emergency logistics scheduling with multiple supply-demand points based on grey interval. Journal of Safety Science and Resilience.

[103] Xu, Xinrun, Lian, Zhanbiao, Wu, Yurong, Lv, Manying, Ding, Zhiming, Yan, Jin, Jiang, Shan. (2024). A Multi-constraint and Multi-objective Allocation Model for Emergency Rescue in IoT Environment. 2024 IEEE International Symposium on Circuits and Systems (ISCAS).

[104] Lian, Zhanbiao, Lv, Manying, Xu, Xinrun, Ding, Zhiming, Zhu, Meiling, Wu, Yurong, Yan, Jin. (2024). Cloud-Edge Collaborative Continual Adaptation for ITS Object Detection. International Conference on Spatial Data and Intelligence.

[105] Jiang, Shan, Ding, Zhiming, Zhu, Meiling, Yan, Jin, Xu, Xinrun. (2021). Graph wavelet convolutional neural network for spatiotemporal graph modeling. Journal of Software.

[106] Xu, Xinrun, Zhang, Qiuhong, Yang, Jianwen, Lian, Zhanbiao, Yan, Jin, Ding, Zhiming, Jiang, Shan. (2025). High-Quality Pseudo-Label Generation Based on Visual Prompt Assisted Cloud Model Update. arXiv preprint arXiv:2504.00526.

---