The Research And Application Of Deep Learning In The Field Of Automatic Driving

Automobile brings convenience to modern society,but also brings traffic accidents.Automatic driving vehicle can solve this problem.The most important part of automatic driving system is high precision and fast perception technology.The perception technology based on computer vision is low cost,simple semantic and highly reliable,and target detection is the main way to realize the perception.Deep learning target detection has some advantages in detection speed and accuracy,but there are also some problems,such as easy to ignore the detection of small size target,fuzzy image is difficult to detect accurately.In view of the above problems,this paper studies object detection based on deep learning.The main work and innovation of this paper are as follows。(1)The improvement of data sets.At present,most of the data sets used in the field of traffic sign recognition are from Europe and the United States,which are different from domestic traffic signs.In order to improve the applicability of the algorithm in the actual scene,according to the national road traffic signs standard,the traffic signs detection data set is improved and established by combining the standard picture with the real picture.(2)Image deblurring processing.In actual driving,because of vehicle shaking and various weather reasons,the images obtained by on-board camera equipment are often fuzzy,which greatly reduces the accuracy of image recognition and increases the probability of danger.In this paper,through the method of enhancing data sets,the detection network is trained to adapt to the changeable weather environment,and the robustness of the detection algorithm is enhanced,so as to avoid recognition errors due to fuzzy problems.(3)Small target detection can detect traffic signs early and make a safe decision in advance,so the recognition system needs to identify the far smaller traffic signs.In fact,the performance of vehicle vision sensor will lead to too few traffic sign pixels in the image.The resolution of the actual sensor is different from that of the training image,so it is necessary to cut or stretch the image to detect the target.After the image is cut or stretched,there will be a certain distortion,resulting in the remote small-scale signs can not be detected.In this paper,based on the multi-scale traffic sign detection network of DetNet network and Yolo network,a multi-scale target detection algorithm is designed to ensure the detection accuracy of small targets by maintaining the size of output feature map.