Design And Implementation Of Real-time Detection Technology For Target Vehicles Under Edge Computing Environment

With the rapid popularization and development of smart cities,a large number of public transport cameras are installed at road traffic intersections.The traffic system captures road conditions in real time by processing the video from surveillance cameras,discovering abnormalities in time,and maintaining public transport safety.Among them,the real-time detection and identification of target vehicles(such as missing and hit-and-run vehicles)is an important application in the intelligent transportation system.Real-time video analysis applications require higher hardware resources due to the deployment of complex image processing models and low latency requirements.Existing researches on target vehicle detection focus on improving the accuracy of image recognition.Most of the models are trained on small-scale offline datasets based on servers with high computing power,while ignoring the real-time problem of video analysis.Today,the explosion of data volume brought by large-scale Io T devices has caused problems related to high bandwidth,high latency,and low privacy in traditional centralized cloud server data processing.As a result,the concept of edge computing came into being.Edge computing enables data to be processed locally on edge servers close to the data source,which reduces the network latency of data transmission and is suitable for real-time data processing scenarios.However,the current related research still has some deficiencies in realizing real-time target vehicle detection on edge devices.on the one hand,few studies consider the realization of real-time target vehicle detection applications on edge devices with limited computing power.On the other hand,the deployment of complex vehicle recognition models on edge devices with limited resources will also bring higher computing costs.When the task volume of vehicles to be recognized at the intersection is large,a large network delay will also occur,making the task unable to be processed in time.To this end,this paper proposes a dynamic multi-granularity vehicle feature recognition strategy based on reinforcement learning,which reduces the computing power consumption of edge devices through simple models under high task load.This paper analyzes the whole process of target vehicle detection and recognition in detail,and divides the vehicle feature recognition model into five levels through the granularity classification of vehicle features.Different models have different complexity.In this paper,the complexity of the vehicle feature recognition model is adjusted to dynamically reduce the resource occupancy,and the scheduling between the models is based on reinforcement learning.Considering the dynamic vehicle tasks generated by real-time traffic,this paper adaptively matches the vehicle tasks and models of different complexity based on the dynamic task volume of the intersection,saves the time of task processing,and realizes real-time target vehicle retrieval.The Actor-Critic algorithm framework performs the training of the policy network.Finally,this paper conducts experiments and tests based on real vehicle traffic data.The experiment uses the vehicle traffic dataset in Shenzhen to simulate the real-time road conditions of the traffic intersection,and trains and evaluates the reinforcement learning model on the dataset.Compared with other baseline algorithms,the performance of the proposed model is more advantageous.In summary,this paper introduces reinforcement learning technology into the real-time target vehicle detection problem,realizes dynamic vehicle feature recognition and matching by building a multi-granularity vehicle feature recognition model and reinforcement learning framework,and reasonably utilizes the limited computing resources of edge devices according to the amount of tasks to achieve real-time vehicle task processing.