Research Of Multiobject Detection And Tracking Via Deep Convolutional Neural Network

Nowadays,with the development of information technology,autonomous driving,as a product of a new generation of information technology,integrates computer vision,highperformance computing,pattern recognition,intelligent control,and other technologies to achieve intelligent vehicle driving,which is currently one of the hot research areas.As a key step of autonomous driving,environmental perception is the fundamental part in the autonomous driving system,a decisive factor in ensuring the safety of autonomous driving.Specifically,as a key technology,multitarget detection and tracking can perceive targets,and maintain stable tracking,thereby improving the driving safety of the vehicle.In this thesis,aiming at the multiobject detection and tracking technology in the autonomous driving environment,an improved Deep Sort multiobject detection and tracking model is proposed,by integrating the idea of convolutional neural network,multiscale learning,gated structure,and attention mechanism.The main contents are listed as follows:1.An attentive gated Res2Net-based YOLOv4-tiny(AGYOLOv4-tiny)object detection model is proposed.Firstly,the multi-scale residual module(i.e.,Res2Net)is introduced to optimize the feedforward network of YOLOv4-tiny,thus forming a Res2Net-based YOLOv4-tiny model to enhance the multiscale reprerentation ability of the original YOLOv4-tiny.Secondly,aiming at the uncontrollable problem of information interaction in Res2 Net,the gated Res2Net-based YOLOv4-tiny model is proposed in conjunction with Gated Res2 Net module.After that,by combining with the attention mechanism,an AGYOLOv4-tiny model is further sconstructed to enhance the useful target information.The experimental results on the public object detection data set verify the effectiveness of the proposed model.2.The improved Deep Sort multiobject tracking model(IMDeep Sort)is developed.Considering the large number of parameters and the slow detection speed of the original Deep Sort multiobject detection and tracking algorithm,the AGYOLOv4-tiny model proposed in this paper is considered as the backbone of the detection network Furthermore,the central loss function is applied to optimize the deep appearance feature extraction module in the tracking network,and then,the complete intersection over union(CIo U)is used to optimize the matching process.Experiments are conducted on the public multiobject tracking KITTI dataset,whose results prove the advantages of the proposed model.