Design And Implementation Of Distributed Real-time Video Target Tracking System Based On Stream Computing

Intelligent unmanned equipment plays an indispensable role in modern military and civil fields because of its low cost,high mobility and zero casualties.Personnel in equipment monitoring,organize rescue work,need by intelligent target tracking system was carried out on the acquisition of video image transmission and use the target detection and tracking of the depth of the calculation and analysis of learning algorithm model,to provide decision basis for the next action,because of the uav,the car carry a limited hardware equipment performance,When the amount of data calculation is large,the system can not run normally due to the lack of single computing power,and the distributed target tracking system based on batch or micro-batch processing will produce batch processing delay and can not be applied to the scene with high timeliness requirements.Based on distributed stream processing technology in view of this,this article puts forward a kind of video target tracking real-time computing framework,first of all video research to improve the multiple target tracking algorithm model,realize the video target tracking speed and precision of the balance,then design and realize the video streaming data message aggregation center,complete video streaming data in a cluster of distributed transmission and storage,Then the distributed video stream processing strategy is proposed,and the effective operation of the deep learning model in the distributed stream processing environment is realized from the steps of target tracking calculation such as data input,processing,model loading and model prediction,and finally the distributed video target real-time tracking system based on the stream computation is realized.The main research contents are as follows:1.Efficiency improved multi-target tracking algorithm.In view of the system’s demand for efficient tracking of video human targets,the YOLOV4 target detection algorithm and DEPSORT target tracking algorithm were tuned in this thesis,which improved the tracking stability while reducing the calculation amount of the model.By organically combining the input and output structures of the two algorithm models,the multi-target tracking algorithm carried by the system in this thesis is realized,and the computational efficiency of target tracking is improved on the basis of meeting the accuracy requirements of the system.2.Video streaming data message aggregation center.In order to realize the efficient transmission,storage and calculation of video data in the distributed streaming computing environment,this thesis proposes a video streaming data aggregation center based on Kafka to complete the data processing and calculation of decoupling.At the data receiving end,the video data is serialized and the partitioned storage strategy is designed to complete the distributed storage and transmission of video streams.At the data consumption end,a consumption strategy based on video segmentation is designed from the point of view of data parallelism,which enables the system to parallelize the data stream for real-time calculation of target tracking,which makes the system highly available.3.Distributed video streaming strategy.To realize the distributed operation of the target tracking algorithm,this thesis based on the Flink distributed stream processing framework,from the data input,processing,loading,model prediction steps such as target tracking calculation,designs and realizes a kind of target tracking task handling strategy under stream computing environment and the corresponding module,finally realizes the cluster for video object tracking tasks distributed real-time operation.