Object Detection And Depth Estimation Based On Deep Learning In Traffic Scene

As people have higher requirements for efficient and safe transportation,it is particularly important to handle target detection and distance estimation in transportation.At present,there are problems in target detection and distance estimation,such as low target recognition accuracy,inaccurate position positioning,sparse radar point cloud data,and difficulty in multi-sensor data fusion.In response to these problems,this thesis uses deep learning technology to improve the mainstream target detection and image depth estimation(distance from the object to the camera in the image)algorithms,and applies them to pedestrian and vehicle recognition and positioning in traffic scenarios.The main work is as follows:Aiming at object detection in traffic scenes,this thesis is based on the YOLOv3 to detect 6 types of targets such as vehicles and pedestrians on the road.For generating the anchor box of the object,this paper adopts the AIoU clustering distance considering the aspect ratio information.Compared with a priori boxes generated by traditional clustering methods,the average IoU index value of a priori boxes generated based on AIoU clustering is slightly higher and the average detection accuracy of the model is improved by 1%.In addition,this article uses LIoU as the new objection positioning loss.The experiments show that the objection detection accuracy of the trained Model for 6 types of targets with the new loss function is increased from 70.6%to 79.8%,and the convergence speed of the Model is higher than that of the prevIoUs model.About image depth estimation,this thesis proposes a loss function that combines supervised loss and self-supervised loss,and completes the model training on the Kitti raw data set based on the Monodepth2 network method.When the estimation error threshold is<1.25,the depth estimation accuracy of the Monodepth2 depth estimation model trained based on the hybrid supervision loss function is 87.0%,which is 3%higher than the model trained by the self-supervised loss.The position of the object in the image predicted by YOLOv3 dtection Model and the depth of the image estimated by the Monodepth2 image depth estimation Model.are obtained respectively.At last,this thesis combines object detection task with the image depth estimation task and obtains the position of the object relative to the camera and the distance information of the area where the object is located at the same time through the unified coordinate system,which better solves the problem of identifying and positioning targets such as vehicles and pedestrians in traffic scenes.