Design And Implementation Of Pedestrian-vehicle Detection And Tracking Algorithm For Driver Assistance System

With the continuous development of artificial intelligence,driving assistance has become one of the development directions of automobile intelligence,and the pedestrian-vehicle detection and tracking algorithm is an absolutely necessary part of it.In the last few years,detection and tracking algorithm based on deep learning has become a major research direction,but there are still some problems such as low detection accuracy,frequent changes of tracking target ID,and limited deployment platform.Centering on these problems,this dissertation presents a pedestrian-vehicle detection and tracking algorithm for driving assistance and deploys it to the embedded platform to achieve accurate real-time detection and tracking of pedestrian-vehicle.In allusion to the problem of low detection precision of pedestrian-vehicle targets in complex traffic scenes,select and improve the YOLOv4 network to detect pedestrians and vehicles,and learn from Dense Net’s idea of cross-layer connection,and introduce the depthwise separable convolution to improve the single-path quintic convolution and SPP and adjacent convolution module in the network respectively,and cross-layer fusion modules Dense-Dw and Dense-SPP are designed.What’s more,a gentler activation function SMish is proposed,which can fully integrate network features and reduce references.The experimental results of the Dense-SMish-YOLOv4 model designed in this dissertation and other detection models on KITTI road dataset show that the improved algorithm is consistent with YOLOv4 in detection speed,while the m AP has increased by 2.2%,which improves the detection ability of pedestrian-vehicle on the whole.For the problem of frequent ID transformations caused by occlusion in the process of target tracking,the detection algorithm is replaced with Dense-SMish-YOLOv4 algorithm based on the Deep Sort algorithm.Kalman filter algorithm,Hungarian algorithm and cascade matching methods are used to complete the task of track association.For the severely occluded situation,the adjacent tracking algorithm is used to associate the lost targets.Experiments on MOT16 dataset show that the tracking accuracy of the improved algorithm reaches 64.6%,which is better than other comparison algorithms.Besides,the improved algorithm can reduce the number of ID transformations to a certain extent,which has certain reference value for solving the occlusion problem in the process of pedestrian-vehicle tracking.In allusion to the problem that the target detection and tracking model in this dissertation has high requirements on the computation platform and cannot directly implement forward reasoning on the embedded platform,the Dense-SMish-YOLOv4 detection model is lightened: take the place of the backbone network with a lightweight network Ghost Net and on this basis,replace standard convolution in the Neck network with depthwise separable convolution.The obtained lightweight model YOLOv4-Light is used as the detection module of the tracking algorithm to construct a lightweight multi-target tracking algorithm.The trained lightweight detection and tracking models are optimized and accelerated by Tensor RT,the reasoning engine of NVIDIA Jetson TX2 platform.Finally,the optimized model is compared with other models on TX2.The experiments show that: although the lightweight model has little real-time advantage,it has outstanding performance in the accuracy of detection and tracking,and can be of great help to the practical application of driving assistance.