Deep Learning Based Location And Recognition Of Traffic Participants

Location and recognition of traffic participants is a key technology in the intelligent transportation system.At present,great achievements have been made in location and recognition for traffic participants in some specific scenes.However,in complex road monitoring scenes,affected by the size,clarity,shape and other factors of the targets,existing location and recognition algorithms are not effective,and it is difficult to achieve a good balance between speed and accuracy.The widespread use of deep learning has made a transformative impact on the development of computer vision,and many well-established location and recognition algorithms are proposed based on deep learning technique.Aiming at the location and recognition tasks of traffic participants in the vehicle-road coordination system,new traffic participant location and recognition algorithms are proposed based on some advanced deep models of object detection and image classification.Experimental results show that the proposed models can significantly improve the accuracy and meet the speed requirement.The contributions are as follow.We constructed new datasets.Images of real roads in different scenes and different time periods are collected by surveillance cameras placed in multiple bayonets.Based on the collected images,and a dataset for object detection are constructed by screening,labeling and other steps.Setting target boxes separated from the background in the surveillance images,screening and labeling again,the classification dataset is constructed.For the object detection task,the YOLOv5 s model is improved to give a new deep network model,called D-YOLO.The specific innovations include: to make the model having better performance in detecting small targets and stabilize the detection accuracy,the detection scales are added to the YOLOv5 s and the depth of the network is increased;the CSPSPP structure is introduced,and the feature is divided into two components as which is done in CSP-Net,to fully utilize the features and reduce the amount of computation;the Focal Loss is applied to the confidence loss and the classification loss to overcome sample imbalance;the Distribution Focal loss and the Focal CIOU+ loss are added to promote the accuracy of prediction of bounding boxes and center points.Training and testing are carried out on the self-built data set.The experimental results show that,our detection model D-YOLO perform well in location of traffic participants,In particular,D-YOLO can achieve 91.4% m AP value and 81.7% Recall value on the dataset of complex scenes For the image classification task,a new classification model,called T-Res Net is proposed based on Res Net18.Specific innovations include: the number of residual blocks in the network is reduced and feature fusion blocks are introduced,to enhance the expressiveness of features and control the depth of the network,make it suitable for large samples;depthwise separable convolution in the residual blocks are used to replace some convolutional layers,so that the amount of computation is reduced,and the classification process can be done at less time cost;some shallow BN layers are removed,so that the network can better capture the feature distribution;The Combined Margin Loss is introduced in learning features in the angle space,so that the gap intra-cluster is reduced and gap inter-classes is enlarged.Training and testing are carried out on the self-built data set.The experimental results show that,our classification model T-Res Net perform well in recognition of traffic participants,with higher accuracy,at required realtime speed.In particular,T-Res Net can achieve 99.5% accuracy and 9.24 ms speed for the seven-category testset.The final location and recognition model of traffic participants achieves an accuracy of 81.3%.