Research On Preceding Vehicle Behavior Recognition Method In Complex Environment

With the development and progress of society,the intelligent transportation system has developed rapidly.Among them,intelligent vehicles,as the main traffic participants,have attracted more attention by researchers.Perception of environment is an important prerequisite for intelligent vehicles to achieve driverless.As a key link,preceding vehicle behavior recognition receives sensor information,identifies the movement of surrounding vehicles,helps the intelligent vehicle make decisions,and plays a key role in driverless driving.This thesis focuses on the types,states,and behaviors of preceding vehicles for intelligent vehicles in complex environments as specific research objects,identifies preceding vehicle behaviors,and verifies system performance through data collected in natural environments.First,the recognition of preceding vehicles in complex environment is based on convolutional neural network.In order to meet the real-time and accurate requirements of vehicle target detection,tiny-YOLOv3 algorithm is used to solve the problem of low detection accuracy of smaller vehicle targets by increasing the network width and detection scale,so as to realize real-time vehicle detection in complex environments.The experimental results show that compared with the traditional tiny-YOLOv3 method,the improved tiny-YOLOv3 meets the real-time requirements,and the network model is smaller and the target detection accuracy is higher,which can effectively identify vehicle targets in complex scenes.Secondly,based on Kernerlized Correlation Filter(KCF),combined with Kalman filtering and Hungarian matching algorithm for long-term tracking of multi-target vehicles in complex environments.In order to overcome the problem of tracking instability when the target is blocked in moving vehicle tracking,the Kalman filter is used to predict the position of the occluded vehicle,and Hungarian matching algorithm is used to match the detection and tracking targets,so as to realize the long-term multi-target tracking of the preceding vehicle.The experimental results show that the combined Kalman filter and KCF method can achieve accurate tracking of vehicle targets in complex environments.Finally,the behavior recognition of preceding vehicles is based on Long Short-Term Memory(LSTM).In order to solve the problems of low accuracy,large delay and single recognition object of vehicle behavior recognition,this thesis uses the results of vehicle detection and tracking to obtain the size of the vehicle motion trajectory and the size of vehicle target detection box.The above motion characteristics are used as the basic characteristics of vehicle behavior classification,and input the LSTM network to establish a time series model,and the vehicle behavior recognition is completed through the output prediction score.The experimental results show that the average classification accuracy of the vehicle behavior recognition model trained by this method is 92%,which can effectively identify multiple behaviors of multiple vehicles in front,and has more advantages than the existing target behavior recognition algorithm.The purpose of this thesis is to study the behavior recognition methods of preceding vehicles for intelligent vehicle,and the results may be a reference for the environmental perception and intelligent decision-making part of intelligent vehicles.