Research On Automatic Driving Object Detection Based On Deep Learning

In recent years,with the rapid developments of artificial intelligence technology,the world's major technology companies have strategically shifted their attention to the field of automatic driving,which pushed the research of automatic driving to a climax.Object detection is one of the core technologies in the field of automatic driving.It is an important way for automatic driving to perceive the environment and affect the control decision of vehicles.Pedestrian,vehicle and non-motor vehicle are the key objects of vehicle driving process.Therefore,according to KITTI automatic driving data set,this paper combines the detection objects into three categories of car,pedestrian and cyclist,and then designs the method applicable to automatic driving object detection.The environment of automatic driving object detection is very complex.There are many problems such as large scale changes,dense objects and serious occlusion.The traditional object detection methods are difficult to meet the requirements of detection accuracy.Therefore,this paper selects two object detection algorithm base on deep learning,they are Faster R-CNN,a two-stage object detection algorithm,and SSD(Single Shot Multibox Detector),a one-stage object detection algorithm,to improve the two algorithms combined with the experimental scene and the characteristics of the algorithm itself,and to verify the performance of the improved algorithm through experiments.For the two-stage object detection algorithm Faster R-CNN,the input size of the network is adjusted according to the resolution of data set.Then ROI Align is used to improve the rough pooling of candidate areas in Faster R-CNN and Focal Loss is used to optimize the classification part of the loss function to pay more attention to the difficult samples during training.The post-processing method of the detection box is improved to Soft NMS(Soft Non-Maximum Suppression)and finally the improved Faster R-CNN based automatic driving object detection algorithm is obtained.Compared with the original algorithm,the improved algorithm improves the m AP(mean Average Precision)from 0.682 to 0.833 on the test set,but the detection speed is only 9fps.For one-stage object detection algorithm SSD,the input size of the network is adjusted considering the aspect ratio of the data set.Then the backbone of SSD is improved based on FPN(Feature Pyramid Networks)to make the semantic information of low-level features fully expressed through multi-scale feature fusion.The feature extraction ability of irregular objects is improved by using modulated deformable convolution in the backbone and the GHM(Gradient Harmonizing Mechanism)loss function is used to improve the classification loss function and the box regression loss function,the improved SSD based automatic driving object detection algorithm is obtained.The improved algorithm's m AP is improved from 0.65 of the original algorithm to 0.818 on the test set,and the detection speed can reach 29 fps.The improved SSD and Faster R-CNN have greatly improved the detection accuracy and achieved similar detection accuracy,but in terms of detection speed,the improved SSD far exceeds the latter,achieving real-time performance and more suitable for the scene of automatic driving object detection.