PaperlayerView exact revision on arXiv

Source-linked AI summary

Distributed Deep Learning Model for Intelligent Video Surveillance Systems with Edge Computing

Jianguo Chen, Kenli Li, Qingying Deng, Keqin Li, Philip S. Yu

arXiv:1904.06400v1cs.CVcs.LG

TL;DR

Centralized video surveillance faces communication overhead and challenges in synchronizing distributed models and balancing heterogeneous edge workloads. The paper develops DIVS, a multi-layer edge-based distributed deep-learning system with parallel training, parameter synchronization, and dynamic data migration. Experiments report decreasing execution time as edge nodes increase, alongside effective workload balancing and video-analysis handling.

  • Problem

    Video surveillance must handle large video streams while addressing communication overhead, distributed-model synchronization, architecture design, and workload imbalance across edge nodes.

  • Method

    The paper deploys a distributed deep-learning model in a multi-layer edge-computing architecture with task-level and model-level parallelism, parameter updating, and dynamic data migration.

  • Results

    Execution time decreases as edge-node count increases, while dynamic data migration maintains steadier workload balance; the DIVS system efficiently handles video surveillance and analysis tasks.

  • Takeaways & Limitations

    The implemented DIVS system provides flexible, scalable edge training capabilities for distributed video surveillance and analysis.

Abstract

from arXiv · show

In this paper, we propose a Distributed Intelligent Video Surveillance (DIVS) system using Deep Learning (DL) algorithms and deploy it in an edge computing environment. We establish a multi-layer edge computing architecture and a distributed DL training model for the DIVS system. The DIVS system can migrate computing workloads from the network center to network edges to reduce huge network communication overhead and provide low-latency and accurate video analysis solutions. We implement the proposed DIVS system and address the problems of parallel training, model synchronization, and workload balancing. Task-level parallel and model-level parallel training methods are proposed to further accelerate the video analysis process. In addition, we propose a model parameter updating method to achieve model synchronization of the global DL model in a distributed EC environment. Moreover, a dynamic data migration approach is proposed to address the imbalance of workload and computational power of edge nodes. Experimental results showed that the EC architecture can provide elastic and scalable computing power, and the proposed DIVS system can efficiently handle video surveillance and analysis tasks.

I. INTRODUCTION

The paper proposes DIVS, a distributed intelligent video surveillance system that combines deep learning with edge computing. It targets communication overhead, distributed-model synchronization, architecture design, and workload imbalance while providing parallel training and dynamic data migration.

  • Research challenges: The proposed system addresses synchronization of distributed AI models, feasible edge-architecture design, and workload balance across heterogeneous edge nodes.These challenges arise under unbalanced terminal connections and unequal edge-node computing capacities.
  • Motivation and scope: Existing video-surveillance AI primarily analyzes static images, while distributed systems must also handle large-scale streaming data and communication demands.The paper identifies massive monitoring terminals, huge video streams, and network communication overhead as central challenges.
  • Motivation and scope: DIVS combines artificial intelligence and edge computing to move video-analysis workloads from centralized network locations toward network edges.The intended benefits are reduced communication overhead and low-latency analysis solutions.
  • Contributions: A multi-layer edge architecture and distributed deep-learning training model are established for DIVS.The architecture migrates workloads to network edges to reduce communication overhead and support low-latency analysis.
  • Contributions: Task-level and model-level parallel training methods accelerate distributed video analysis.Different deep-learning sub-models run analysis tasks in parallel, while CNN training is further parallelized within edge nodes.
  • Contributions: Model-parameter updating supports global-model synchronization with low communication cost, while dynamic data migration improves workload balance.The migration approach considers both unequal monitor-terminal connections and unequal edge-node computing capacities.

II. RELATED WORK

The paper situates DIVS within distributed AI, video surveillance, and edge-computing research, then describes its multi-layer architecture and edge-deployed distributed learning model. The system processes streaming video at edge nodes while coordinating scheduling, data, resources, and workload balancing.

  • Related work: Prior work spans distributed AI and deep learning in distributed, cloud, fog, and edge-computing environments.These approaches aim to improve performance and scalability across computing settings.
  • Related work: Video-surveillance research has applied CNNs, neural networks, and related models to face recognition, person re-identification, vehicle detection, and traffic monitoring.The cited applications cover several public-safety and transportation tasks.
  • Related work: Distributed, cloud, and edge deployments have been explored to handle large-scale video datasets and improve surveillance performance.Examples include distributed face recognition and edge platforms that offload computing tasks between monitoring devices and edge nodes.
  • System architecture: DIVS combines IoT, AI, and edge computing in a multi-layer platform containing monitoring devices, edge nodes, a cloud server, and a distributed deep-learning model.The platform is intended to provide flexible, scalable computing capabilities and reduce network communication overhead.
  • Distributed processing: The cloud server performs task scheduling, data management, resource allocation, and dataset migration for workload balancing.It monitors training time costs on edge nodes during model training.
  • Distributed processing: Monitoring terminals stream surveillance video to corresponding edge nodes, where streams are divided into frames and task-specific data blocks for deep-learning sub-models.Each sub-model trains on its input dataset and updates local parameters.

IV. IMPLEMENTATION OF DIVS SYSTEM

The implementation accelerates DIVS through task-level and model-level parallel training using CNN and LSTM sub-models for video analytics. CNNs classify vehicles, while LSTMs predict traffic flow from the monitoring streams.

  • Implementation overview: The implementation addresses parallel training, model synchronization, and workload balancing in the DIVS system.Its parallelization methods are intended to accelerate video analysis, while separate methods handle synchronization and data migration.
  • 1) Task-level Parallel Training:: Task-level parallel training deploys multiple deep-learning models with different structures on each edge node for concurrent analysis tasks.The paper gives CNN vehicle classification and LSTM traffic-flow prediction as examples sharing traffic-monitoring video streams.
  • 1) Task-level Parallel Training:: The CNN model extracts vehicles from video frames and classifies the resulting vehicle sub-images.The model is copied and distributed to each edge node for parallel training.
  • 1) Task-level Parallel Training:: The LSTM model predicts traffic flow by learning temporal dependencies in video streams shared with the CNN vehicle-classification model.Its structure contains input, recurrent hidden, and output layers, with gates controlling the memory cell.
  • 1) Task-level Parallel Training:: The LSTM traffic-flow prediction model is copied and distributed across edge nodes for parallel training.The model represents monitoring video streams as time series and uses memory-cell gates to process them.

2) Model-level Parallel Training:

Model-level parallel training accelerates distributed CNN sub-models by parallelizing convolutional areas and neuron computations on edge-node multi-core CPUs.

  • Model-level Parallel Training: Model-level parallel training uses edge nodes’ multi-core CPUs to accelerate training within each distributed deep-learning sub-model.The paper illustrates parallelization for convolutional and fully connected layers using CNNs.
  • Convolutional Layer: Convolutional inputs are partitioned into areas processed independently with the filter matrix to produce feature-map elements.Because separate tasks access different input areas without updating them, the convolutional computations have no data dependency.
  • Convolutional Layer: The convolutional algorithm computes PConv = Da × Ha × Wa parallel tasks, extracts an area for each task, and appends each convolution result to the feature map.The inputs are the training-video frame matrix X and the current CNN filter matrix F.
  • Fully-Connected Layer: Neurons within the same fully connected layer can execute in parallel because they have no logical or data dependency with one another.Dependencies remain between adjacent layers because each neuron connects to all neurons in the preceding layer.
  • Fully-Connected Layer: The fully connected network decomposes hidden-layer and output-layer computations into independent sub-tasks, with maximum parallelism equal to the widest layer.The hidden and output layers are represented by m and k parallel sub-tasks, respectively.

B. Weight Parameter Update and Model Synchronization

The synchronization method aggregates local weight sets from heterogeneous edge nodes into a global model after batch training, then redistributes the updated global weights for the next batch.

  • Motivation: Uneven terminal connections and heterogeneous edge-node capacities create different workloads and training speeds, causing synchronization problems during global weight updating.The paper proposes weight updating and model synchronization to address these conditions.
  • Weight Sets: A local weight set contains parameters for a sub-model trained from local video frames and updated on a low-level edge node.Each edge node trains its corresponding sub-model using local data.
  • Weight Sets: A global weight set contains parameters for the entire deep-learning model and is updated on a high-level edge node by collecting local weight sets.The high-level edge node aggregates the local sets from all sub-models.
  • Synchronization Workflow: After batch training, the latest local weight sets are aggregated at a high-level edge node to calculate a new global weight set.The initial global set is shared with low-level edge nodes, which update local sets as they process batch samples.
  • Weight Updating: The global weight set is computed from local weight sets using contributions based on edge-node batch sizes relative to the average batch width.The method defines batch sizes for individual and average edge nodes before calculating each local set’s contribution.
  • Synchronization Workflow: The updated global weight set is shared with each edge node for the next batch, and the process repeats until the weight set reaches a steady state.The workflow is described as the paper’s weight parameter update and model synchronization procedure.

C. Dynamic Video Data Migration

The Dynamic Data Migration strategy estimates edge-node imbalance from training times and video-frame counts, then matches migrations to balance workloads while limiting unnecessary transfers.

  • DDM computes each edge node’s epoch time as per-frame training time multiplied by its assigned video frames.The migration assessment period is the interval between two migration operations.
  • Nodes below the average training time migrate frames, while slower nodes receive frames when they have available capacity.The sign of T − T_j determines whether a node sends or receives video frames.
  • The strategy formulates workload balance by minimizing the dispersion of post-migration execution times across edge nodes.After migration, each node’s execution time is recalculated from its updated frame count.
  • Migration is triggered only when imbalance reaches the threshold, rather than continuously during every interval.The threshold θ_B determines whether migration amounts and a migration plan are calculated.
  • DDM matches migration requirements and capacities using an offset ξ, preferring whole-dataset moves and minimizing the number of participating edge nodes.Small unmatched migration or immigration quantities are excluded from the corresponding lists.

A. Experimental Setting

The experiments deploy DIVS as a two-level edge-computing system with monitoring terminals, edge servers, a cloud server, and traffic-monitoring video data.

  • The experimental EC system contains 200 monitoring terminals, 35 EC servers, and one cloud server across two edge levels.Thirty first-level edge nodes collect streams on corresponding streets, while five second-level edge nodes form the upper edge layer.
  • The implementation evaluates dynamic data migration using edge-node lists, training-time measurements, and a matching offset ξ.The algorithm outputs data-migration matching results after computing average execution time and migration amounts.
  • Monitoring terminals and edge nodes communicate over a high-speed Gigabit network, using seven days of collected traffic-monitoring videos.

B. Performance Evaluation

DIVS performance improves as edge-node scale increases, while larger task scale does not proportionally double execution time, indicating scalable processing under the evaluated settings.

  • 71.01%: execution time for 200 MTs falls from 6117.22 (s) with 5 edge nodes to 1773.62 (s) with 30 edge nodes.For 50 MTs, the corresponding decline is 61.91%, from 2335.13 (s) to 851.28 (s).
  • Increasing the number of edge nodes continuously decreases total DIVS execution time for a fixed number of monitoring terminals.At 30 edge nodes, execution time is 1773.62 (s) and 851.28 (s) in the reported cases.
  • Increasing analysis-task scale does not double total execution time because each input video stream is transmitted from monitoring terminals to edge nodes only once.The resulting communication reuse saves data-communication delay across multiple tasks.
  • The execution-time benefits of multiple tasks become more pronounced as the number of monitoring terminals increases.The experiments compare models with similar computational complexity.

C. Data Communication and Workload Balance

The DDM strategy maintains steadier workload balance as monitoring-terminal and edge-system scales increase, while accepting a compromise in data communication cost.

  • DDM provides the most significant workload balance with a compromise data communication cost in most evaluated cases.
  • With DDM, workload balance remains steady as the number of monitoring terminals increases.The evaluation varies monitoring terminals from 30 to 200 and edge nodes from 5 to 30.
  • With DDM, workload balance also remains steady as the EC system scale increases.
  • Without DDM, lower data communication cost in most cases accompanies seriously unbalanced edge-node workloads, longer synchronization waiting, and more total execution time.

VI. CONCLUSIONS

The DIVS system combines multi-layer edge computing with distributed deep learning to support scalable training and efficient video surveillance analysis. It addresses parallel training, global model synchronization, and workload imbalance among edge nodes.

  • The DIVS system is deployed on a multi-layer edge computing architecture to provide flexible and scalable training capabilities.
  • Two parallelization aspects improve the throughput of the DIVS system, while model parameter updating synchronizes the global deep learning model.
  • Dynamic data migration addresses workload and computational power imbalance among edge nodes.
  • The DIVS system can efficiently handle video surveillance and analysis tasks.
Loading 1904.06400v1…