Vehicle Attitude Recognition Based On Deep Learning Method Research And System Implementation

For the vehicle attitude recognition system,its core role is to improve the detection and recognition accuracy of the position and attitude information of road vehicles,so as to judge the driving conditions of the road,so as to further realize the autonomous driving function and ensure driving safety,and improve the driving experience of drivers and passengers.The traditional target detection algorithm is analyzed from the detection principle and process.Such algorithms have problems such as tedious detection process and poor detection accuracy.Therefore,this thesis proposes a vehicle attitude recognition method based on CenterNet based on computer vision recognition technology.In this thesis,the CenterNet algorithm network model based on deep learning is adopted to realize vehicle detection and attitude recognition.According to the research content,this thesis summarizes the problems and solutions into the following four points:(1)To solve the problems of large network model and slow training,the lightweight network architecture EfficientNet is selected to reduce the number of participants.(2)To solve the problem of dimensionality loss caused by dimensionality reduction in feature extraction of vehicle attitude recognition,ECA module was used to replace SE module to optimize the attention mechanism to improve the accuracy of the model.(3)In order to better train the neural network and improve the detection accuracy,the activation function of the backbone feature extraction network was changed to Mish activation function.(4)For the problem of universal lock in attitude representation,the way of network feature regression quaternion is adopted to solve it.Finally,x,y,z and Euler Angle are used to represent the attitude and position information of the vehicle in the picture.Finally,the improved network model is trained and tested.The experimental results show that the improved algorithm can adapt to a variety of different traffic conditions.Regardless of the number of vehicles,the distance and the impact of pedestrians,the detection and attitude recognition of vehicles are realized,and the accuracy rate and recall rate are more than 93%,which reflects the robustness of the improved algorithm.Compared with the previous algorithm model,the size of the improved CenterNet network model is reduced by nearly 13%,and the detection accuracy rate is increased by 1.26%.The overall goal of this thesis is reached,which helps developers to realize more functional operations with the obtained information and better serve consumers.