Research On Classification And Behavior Analysis Of Vehicles In Video Sequences

With the national transportation network becoming more and more perfect,traffic engineering informatization and intelligence have become the mainstream of today’s development,and vehicle classification and vehicle behavior recognition as key technologies among them have also received widespread attention in the industrial and academic fields.In the process of vehicle classification,factors such as mutual occlusion and weather conditions affect the accuracy and stability of the results.When analyzing the behavior of vehicles,the robustness and accuracy of the results often depend on the angle and quality of the video source.These all put forward higher requirements for vehicle classification and behavior recognition technology.Therefore,this paper studies the vehicle classification algorithm and vehicle behavior recognition algorithm,the main contributions are as follows:Aiming at the vehicle classification problem in the video sequence,this paper proposes a vehicle classification method based on the improved YOLOv3 network.Based on the YOLOv3 object detection network,drawing on the design ideas of densely connected convolutional network,the proposed method replaces the residual layer in the original network with a dense convolutional module.And the method adjusts and improves the design structure of the network and the minimum resolution scale of the network features.Multi-scale predictions of vehicle characteristics are made with reference to the Pyramid network model.Using the Softmax function as a classifier,the vehicle features fused by the convolutional layer and the corresponding scale dense convolution module are classified,and the bounding box of the vehicle is labeled based on the anchor box mechanism.According to the classification results of single-frame images,a object matching tracking function is designed to perform continuous and stable detection of vehicles in the video sequence.Aiming at the problem of vehicle behavior analysis in video sequence,a vehicle behavior recognition method based on lane feature fusion is proposed.Based on the lane-line-related data and information,an interpolation formula is established using the improved Robinson operator to obtain the optimal gradient amplitude to extract the lane line edges through image preprocessing,and then the LSD detection algorithm is used to implement the lane line related detection.According to the results of the lane line detection,the lane line feature points are extracted through a sliding window,and the cubic spline interpolation mathematical model is used to complete the lane line feature points fitting.The vehicle’s movement state is analyzed according to the two parameters of the curvature radius and the direction of the lane line,and the deviation state of the vehicle is identified by combining the relative position of the vehicle center and the lane line center.In the test of the BDD100 K public data set,the average accuracy of the vehicle classification algorithm is 93.18%,and the accuracy of vehicle behavior recognition is 93.04%.Experimental results show that the vehicle classification method in this paper can effectively distinguish vehicle types in the field of view and has high accuracy.At the same time,the vehicle behavior recognition method can accurately identify the vehicle’s motion state and deviation state and has high robustness.Both methods have high practical application value.