Reserarch Of Vision Based Vehicle Smart Detection Technology

Autonomous driving system can improve the driving safety according to the information of the environment,and assist the driver to control the vehicle.Therefore,in recent years,with the development of sensor technology,detection and control algorithm,autonomous driving system technology has received extensive attention and research.One of the important foundations of environmental perception,behavior estimation and control decision-making is object detection algorithm.At present,many vision-based 2D object detection algorithms can not accurately provide vehicle pose and size information,although the detection speed is fast.Most 3D vehicle target detection methods are based on the depth information provided by LiDAR,which means that the cost of the devices is high.Visual-based 3D target detection can provide a cost-effective 3D object detection scheme than LiDAR.However,many vision-based 3D vehicle object detection methods require complex neural networks and prior models,which make it difficult to deploy on small platforms or achieve faster detection speed.In this paper,a monocular 3D object detection method is proposed,which uses the vehicle 2D bounding box and observation angle of vehicle to obtain the vehicle 3D bounding box.The proposed method can use the existing 2D object detection network,and the observation angle estimation part is independent of the 2D object detection network,which reduces the requirement of dataset annotation information,and also reduces the difficulty of network design,debugging and training.The proposed algorithm does not need complex network to estimate 7 or 9 Degree of freedom(DoF)parameters representing the vehicle pose.Based only on 2D bounding box,observation angle and vehicle type,vehicle 3D bounding box can be obtained,which has less computation and can be deployed on low performance or mobile platform.In addition,the change of camera intrinsic and extrinsic parameters does not mean that the whole network need to be rebuilt or retrained,which facilitates the modification and transplantation of the algorithm.This paper first evaluates the accuracy of vehicle object detection using YOLOv3-tiny network.In order to match the network with hardware platform,the network depth is deepened.Backbone of the original network is changed to residual network,which could make full use of GPU resources and improves the accuracy of 2D object detection.In order to get the observation angle,this paper uses the residual network and trains it on the processed KITTI dataset,so that the network can output the observation angle information of the vehicle directly according to the vehicle picture.After getting the 2D bounding box and observation angle,this paper uses the conventional computer vision method to solve the equations with prior information such as camera intrinsic and extrinsic,and obtains the 3D position information and the 3D boundaing box of the vehicle.The experimental results on KITTI test dateset show that the AP50 of Easy mode,AP70,AOS and AP3 D have reached 95.1%,77.8%,76.9% and 1.8%.Although there is still a gap with the most advanced object detection accuracy,the requirement of hardware is greatly reduced.The cost of GPU RAM is less than 1GB.In order to apply and visualize the results of vehicle object detection,based on the 2D or 3D information,this paper realized the transformation of road environment to bird's-eye view and lane line display.The whole system can give the position,orientation and lane line of the target vehicle in bird's-eye view map,and the 2D/3D bounding box,lane line and object matching results of the vehicle in the origional image.