Research On Vehicle And Pedestrian Recognition Methods In Vehicle Visual Environment Perception

With the development of my country’s economy,the number of cars has continued to increase,and traffic accidents have also increased year by year.In order to reduce the casualties and economic losses caused by traffic accidents,relevant departments have formulated rules to require operating passenger and trucks to have active safety functions,with vehicle and pedestrian detection as its main Technology,vehicle and pedestrian detection in the driving environment are also indispensable components of smart cars.Compared with other sensors,the camera has more advantages,so vision-based vehicle and pedestrian detection has become the first choice of major auto manufacturers.At present,due to the limitations of hardware performance under the vehicle platform,the detection algorithms at this stage are difficult to achieve the combination of accuracy and speed,so it is of great significance to improve the accuracy and speed of vehicle and pedestrian detection under the vehicle platform.This research focuses on the tasks of vehicle and pedestrian detection in the vehicle environment:(1)Summarize the basic knowledge of convolutional neural networks,summarize and analyze the principles of existing target detection algorithms and the evaluation indicators used in the field of target detection,and provide a theoretical basis for the vehicle and pedestrian detection algorithms in this article.Taking detection accuracy and detection speed as indicators,the existing target detection algorithm is difficult to realize the task of vehicle visual environment perception under the embedded platform.For the vehicle platform,a Fast Vehicle and Pedestrian Detection(FVPD)model is proposed.(2)Through the analysis of the driving environment,the difficulty of vehicle visual environment perception in the external environment and the deployment of the vehicle platform is summarized.Select 185363 images from driving videos to make a data set in the PASCAL VOC data set format,and divide the training set,validation set and test set according to the ratio of 60%,20%,and 20%.Design FVPD model input and enhance the generalization ability of the model through image processing algorithms.Draw lessons from the idea of one-stage target detection algorithm,build FVPD model based on deep separable convolution,residual network fusion convolution block attention module,spatial pyramid pooling,path aggregation network and adaptive spatial feature fusion,and adopt convolutional neural The network directly predicts the target information.(3)The K-means++ clustering algorithm is improved by using the intersection ratio as the distance formula,and nine groups of anchor boxes of different sizes are obtained from the data set.In order to improve the detection of mutual occlusion and nearby targets in the FVPD model,CIo U and cross entropy are used as the loss function of the FVPD model,and the redundant information is filtered by introducing a non-maximum suppression algorithm of center distance.Formulate FVPD model training strategy,build and train with the help of Tensorflow framework,and get model weights.In the laboratory environment,the FVPD model is analyzed and verified from many aspects.(4)For the vehicle deployment environment,the Jetson Xavier NX module is selected as the FVPD model deployment chip,and the embedded system hardware platform deployment plan is designed.The model deployment is modeled on the laboratory software environment.The detection speed is only 1.56 frames/sec,which cannot meet the actual vehicle visual environment perception task.Optimize the FVPD model deployment plan,use the inference engine Tensor RT to deploy in a half-precision manner in the C++ environment.