Study On Vehicle Object Detection Technology Based On Embedded Platform

At present,the rapid development of artificial intelligence technology has accelerated the reform of the automotive industry,and the intelligentization of automobiles has ushered in new development opportunities.The development of intelligent vehicles is of great significance for solving traffic congestion and promoting urban digital construction.In this paper,we carry out research on the in-vehicle environment perception method based on the combination of deep learning-based object detection method and stereo vision to lay the foundation for intelligent vehicle decision-making and control.First,in order to meet the real-time and accuracy of the object detection algorithm on the embedded platform,we compare and analyze different object detection methods,and finally select Yolov3-tiny as the basic framework for object detection in traffic scenarios,and construct Yolov3-tiny’s object detection system.On this basis,we analyze the shortcomings of the Yolov3-tiny model applied to the object detection of traffic scene through experiments.Then,aiming at the problem of missing detection and re-detection of Yolov3-tiny model,we improve the structure of Yolov3-tiny network to increase the detection scale and the depth of the backbone network.And we optimize the network model parameters,and use K-means clustering algorithm to cluster the width and height of the anchor box for the objects in the dataset,and determine the optimal number of anchor boxes through comparative analysis.After training,we get a robust-yolov3-tiny model that detects five types of objects: bicycle,bus,car,motorbike,and person.Under the real-time requirements,the model improves the detection accuracy and enhances the detection effect on small objects.Finally,based on the robust-yolov3-tiny model detection,we merge it with stereo vision technology to correct the size of the object detection frame to further increase the IOU value of the detection frame and the real frame.We use the Kalman filter object tracking algorithm to reduce the missed detection rate,thereby improving the stability of the object detection frame.On this basis,we separately calculate the longitudinal distance and lateral distance of the detected object relative to the reference point,and estimate their motion status.