Research On Vehicle Detection And Tracking Based On Heterogeneous Computing

With the great improvement of material life,the number of vehicles has increased rapidly,and the traditional form of vehicle transportation is failing to meet people's needs for travelling.As a novel way to solve traffic safety problems,automotive autonomous driving technology has become a new research issue in many organizations.Autonomous vehicle driving technology aims to enable the car to automatically complete the perception of the environment and the planning and control the movement of the vehicle without human intervention.Vehicle detection and tracking is an important part of vehicle autonomous driving technology,which has a good application prospect in vehicle assisted driving,danger alarm,activate safety and other aspects.According to the high accuracy,robustness and realtime requirements in the field of environment perception of vehicle detection and tracking,heterogeneous computing platforms have the ability to allocate workloads according to the computational characteristics of different computing resources,and perform parallel computing processes.Accelerated and optimized operating environment advantages,heterogeneous computing platforms can achieve high complexity vehicle detection and tracking solutions while meeting real-time requirements.Based on this,this paper analyzes and summarizes the image features and classification methods corresponding to the target scene,combines the computational characteristics of the mainstream heterogeneous computing platform,implements multiple schemes and compares and analyzes in various data sets,and completes the actual deployment.The main tasks are as follows:1)For the digital image technology involved in the real vehicle forward detection and tracking scene,the digital image features and basic principles of the detected objects in various actual working conditions are discussed.The features of the model are determined through comparative experiments.2)Theoretically deduced the mathematical principles of support vector machine and deep neural network model used in this thesis,then analyzed and discussing the target detection and tracking methods based on these two different models.3)Combining the previous models and the deep learning framework Tensor Flow involved in the experiment,the heterogeneous computing platform for training models,the integrated technology CUDA,the accelerated library cu DNN,and the model optimization after training are introduced,and the paper is introduced.Use the inference engine Tensor RT to build the inference engine and deserialize the workflow deployed to different platforms.4)Introduced the self-acquisition dataset of the domestic scenario and the public dataset KITTI and GTI of the foreign scenario,and implemented and deployed the SVM,YOLO,YOLOv2,SSD and Faster RCNN models on two heterogeneous computing platforms,and adopted the contrast experiment analyzes and studies the accuracy,robustness and real-time performance of the vehicle detection and tracking algorithm based on the vehicle monocular camera solution.Finally,the YOLOv2 FP16 solution is determined for efficient target detection.The scheme is deployed on the real-vehicle Jetson Xaiver platform.The determined solution of this thesis fusion with millimeter-wave radar information provides the vehicle in front of the egovehicle for the forward collision warning system,the adaptive cruise control system,the lane keeping assist system and the active lane change system.The number and location information is now in use.