Stream Data Target Recognition Algorithm And Application Research Based On Domestic Processor

Recently,with the continuous development of intelligent traffic management technology,it has added a lot of convenience to people's lives.Video surveillance is used as one of the main data sources for traffic assistance management.The video data's growth rate far exceeds the current network bandwidth's growth rate,for this reason implementing efficient processing of video stream data is a current research priority.Although cloud computing enables efficient data processing,the high latency of intermediate data transmission makes it difficult to meet the demand for time-sensitive traffic video streaming data processing.Edge computing-related technologies have emerged as the cloud's calculation power derivative to alleviate these problems to some extent.As an edge device,the domestic Loongson big data machine has been promoted due to its convenience and ease of deployment,but the MIPS architecture-based device has weak support for mainstream deep learning frameworks,making it more difficult to bring relevant image processing algorithms to the desired level.To improve the efficiency of data processing based on domestic equipment and to meet the demand for traffic video stream data processing's timeliness,this thesis carries out layered processing of video stream data,i.e.to achieve the separation of the video stream and text stream processing.The main work of this thesis is as follows:(1)Two lightweight traffic object detection algorithms,LENetDet and YoLite+,are proposed to address the problems.The problems include end-side device resources limited to difficulty deploy large models and the network structure changed to difficulty adjust the weight parameters.For LENet Det,the algorithm uses a feature reuse mechanism to achieve model parameter compression and achieve model convergence without pre-trained weights.For Yo Lite+,the algorithm uses pretrained weights,MobileNet deepwise separable convolution,and feature map negative phase information to achieve YOLOv4's parameter compression,allowing the algorithm model to meet the requirements of the business scenario.To improve the above two algorithms' generalization ability in complex traffic environments,a data enhancement scheme for multi-scene simulation is proposed.(2)To improve video stream data' processing efficiency based on the domestic big data machine,a multi-stage edge computing model is proposed for domestic processors.Firstly,on the end-side,the semantic information extraction of video image scenes is implemented using lightweight target tracking algorithms to transform image data into text stream data.Then,on the side,the text stream data from the previous level is analyzed using machine learning regression algorithms based on the big data machine to predict traffic congestion.The cloud is responsible for model training and dynamic tuning.The scheme can effectively improve the data processing efficiency based on domestic equipment,while also reducing the network bandwidth pressure.(3)To provide users with an intuitive visualization,this thesis integrates the above research,implements a traffic stream data processing recognition prototype system based on a domestic big data machine.Through this system,users can view the current area traffic congestion,and realize the traffic target marking in the picture.