Research On On-board Object Detection Methods Based On Deep Learning In Complex Scenes

In recent years,many internet companies have focused their strategic goals on autonomous vehicles,and set off a worldwide research upsurge of autonomous vehicles technology.In the application scenario of autonomous vehicles,achieving self-driving in complex scenes such as urban roads is more challenging.The key of self-driving is to identify the objects in the driving environments,and the surrounding vehicles and pedestrians which need to be paid attention to in the urban road environment.Because the information of vehicles and pedestrians in driving environment affects the decision-making of autonomous vehicle control.This paper designs a object detection model which is suitable for the vehicle image in the complex road scenes.In order to solve the problem of frequent occlusion and changeable posture in on-board object detection,the deep learning methods based on convolution neural network are more effective than the traditional object detection methods.The single-stage method of SSD is not only faster but also more effective than the two-stage methods or other single-stage methods.After studying the differences of object detection methods,SSD300 model with excellent performance on Pascal VOC was selected as the basic model.This paper analyzes the differences between KITTI and Pascal VOC,proposes a model design method considering pixel proportion of objects,information distribution of objects and aspect ratio of objects.Finally,a detection model called KITTI_SSD650*150 was designed for KITTI.In order to verify the effectiveness of the method used in the process of detection model design,several groups of comparative experiments were designed.The detection performance of SSD300 model and KITTI_SSD300 model on KITTI is compared,which verifies the effectiveness of considering pixel proportion of objects.The performance of KITTI_SSD300 model is respectively compared with KITTI_SSD512 model and KITTI_SSD650*150 model on KITTI,which verifies the effectiveness of considering information distribution of objects and considering aspect ratio of objects.The detection performance(mAP)of the detection model on KITTI increased from 35.12 % of the SSD 300 model to 77.27 % of the KITTI_SSD650*150 model.