The Research On Vehicle Behavior Recognition And Detection Based On Deep Learning

In recent years,with the increasing transformation of the automotive industry and a spurt of progress in artificial intelligence,unmanned driving technology has been widely promoted and applied.Unmanned environment-aware technology is an indispensable technology for driving vehicle safety and effectiveness.The visual environment data around the vehicle are recorded by the vehicle-mounted camera,which entails that the surrounding vehicles or obstacles are identified to perform corresponding operations.The accurate identification of the behavior of moving vehicles ahead is the foreground for the realization of driverless technology.Simultaneously,vehicle behavior detection plays an irreplaceable role in intelligent traffic management.At present,the traffic violation audit is in a measure faulty as a consequence of overinvestment in human capital,low efficiency,and serious waste of human resources.The video captured by traditional intersection monitoring can only detect traffic violations in a small range,far from the extent of traffic accident prevention and traffic violation detection in full coverage in smart traffic.Most of the vehicles are now equipped with driving recorders or other environmental recording devices in front of the vehicle.As such,the video on the behavior of the vehicles in front of the driving has saw an exponential growth,among which the vehicle's various behaviors are identified and detected in an intelligent and efficient manner,meaning that this ballooning technology is a boon to the unmanned technology and intelligent transportation.In this connection,after the existing vehicle behavior recognition and target behavior detection technology has been fully researched and improved,the research on classification and detection of vehicle dynamic behavior based on deep learning is realized.The data set is collected on the driving vehicle video recorded by equipment such as driving recorders.The research contents include:Aiming at the low precision and large delay of traditional vehicle behavior recognition algorithm,a deep learning algorithm for vehicle behavior dynamic recognition based on long short-term memory was proposed.Firstly,the key frames in the vehicle behavior video were extracted.Secondly,the dual convolutional network was introduced to analyze the feature information of the key frames in parallel,and then the long short-term memory network was used to sequence the extracted characteristic information.Finally,The predicted score of the output determined the behavior of the vehicle.The experimental results show that the proposed algorithm has an accuracy of 95.6%,and the recognition speed of a single video is only 1.72 s.The improved double-convolution algorithm improves the accuracy by 8.02% compared with the ordinary convolutional network.The accuracy of the traditional vehicle behavior recognition algorithm of the data set is 6.36%.Aiming at the problem that traditional manual labeling and positioning vehicle behavior is too time-consuming and labor-intensive,the corresponding target detection and recognition algorithm has poor practicability and low behavioral location detection rate,a vehicle behavior detection algorithm based on dual-stream convolution and two-way long-term and short-term memory network is proposed.The straight-line behavior of the vehicle during driving has occupied a huge proportion,but in actual demand,it pays more attention to the turn of the vehicle in the long video,and the behavior of turning the road down.Therefore,this article will default to the background behavior,the vehicle from the long video.Changing lanes,turning and turning around are set as target behaviors.First,the non-direct behavior proposal in the long video is extracted through the dual-stream convolution network,and then the preliminary proposal is pruned and identified using the two-way long-term and short-term memory network,thereby locating the location and behavior category of the behavior.Based on the data set of this paper,it has obvious advantages compared with the existing target behavior detection algorithm.When the tIoU threshold is 0.5,the mAP reaches 36.3%.