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

As people demand more and more efficiency and safety in traffic travel,it is especially important to solve the problem of target detection and depth estimation in traffic travel.In order to prevent traffic accidents,traveling vehicles must keep a safe distance from pedestrian vehicles in front of them.The current target detection and depth estimation tasks have problems such as low target recognition accuracy,poor position localization accuracy,sparse radar point cloud data and difficult multi-sensor data fusion.In this thesis,based on the deep learning approach,the target detection algorithm and the monocular depth estimation algorithm are combined to output a target detection image with depth information for a given arbitrary single image.The main research of this thesis is as follows:(1)To address the problem of low accuracy of vehicle and pedestrian recognition in traffic scenes,this thesis proposes a target detection algorithm based on YOLOv5.The main framework of this algorithm includes the parts of feature extraction backbone network,feature fusion module,and target prediction.The attention mechanism is added in the feature extraction backbone network part,and the GIo U＿Loss loss function is replaced by CIo U＿Loss loss function.The experimental results show that the recall of the model improves from 90%to 93%after the improvement on the validation set.The proposed network model,which not only improves the prediction accuracy but also effectively reduces the loss,proves the superiority of the model in this thesis.(2)A self-supervised monocular depth estimation algorithm based on channel attention with mainly encoder-decoder architecture is proposed to address the problem of unclear local details in prediction maps for image depth estimation.A structure-aware module and a detail-emphasis module are added to the depth prediction network to capture more contextual information of the scene and emphasize detail features.The superior performance of the proposed method is verified on the KITTI dataset and Make 3D dataset.The accuracy of the improved Monodepth2 monocular depth estimation model based on the improved Monodepth2improves from 98.1%to 98.3%at an estimation error thresholdδ<1.25~3.(3)In this thesis,an algorithm of target detection and depth estimation based on deep learning is proposed.The task of target detection and depth estimation is connected in parallel by supervised learning,and an end-to-end multi-task model of real-time depth estimation and target detection is proposed.