Multi-Object Segmentation And Tracking For Autonomous Driving Scene

In recent years,autonomous driving technology has received more and more attention due to its safety and good application prospects.However,in the autonomous driving scene,there are problems such as multi-scale,geometric deformation,and occlusion of the object,which causes the vehicle perception system to be unable to accurately segment and track the surrounding objects,and further makes the control system unable to obtain the real scene information.In response to the above problems,this paper introduces densely connected feature pyramids,deformable convolutions,and fusion of local information of objects on the framework of multi-object segmentation and tracking algorithms based on convolutional neural networks.Deformation and occlusion issues have been studied,mainly including:(1)Aiming at the multi-scale situation in the autonomous driving scene,this paper introduces the densely connected feature pyramid structure into the image feature extraction stage on the multi-object segmentation and tracking framework based on the convolutional neural network,and shortens the difference between features of different scales.Connecting paths makes the feature map contain both low-level spatial features and high-level semantic features.This paper uses the "detect first and then track" network Track RCNN and spatio-temporal clustering-based network STEm-Seg to conduct experiments.Both algorithms have the ability to detect multi-scale objects after using a densely connected pyramid structure.rise.At the same time,in view of the weak ability of the convolutional neural network to model the geometric deformation of the object,the author introduces the deformable convolution module into the multi-object segmentation and tracking network.The results show that this module can improve the model’s detection of multiple geometrical shapes.ability.(2)Aiming at the occlusion problem that occurs during the driving process of the vehicle or the pedestrian crossing the road,at the same time,this causes the interruption or fragmentation of the tracking trajectory.In this paper,the idea of part and whole is introduced into the multi-object segmentation and tracking framework based on convolutional neural network.The object is divided into several regions,and the features of each region are weighted and added to replace the overall features of the object.This method can make the visible part of the object and the whole characteristic change continuous.The results show that this method can effectively solve the problem of interruption or fragmentation of tracking trajectory caused by mutual occlusion between objects.This thesis has 31 figures,5 tables and 88 references.