A Study On Object Detection For Traffic Scene Based On Improved Faster RCNN

In the field of traffic safety,unmanned driving technology has gradually matured.Using deep learning algorithms to assist in the recognition of pedestrians,vehicles,traffic signs,signal lights and other traffic targets has become the key to achieving unmanned driving.There are still many problems in the application of convolutional neural network to the traffic object detection: 1)the accuracy of multi-scale object detection in complex traffic scenes is not enough,and it is difficult to detect small targets;2)It is difficult to distinguish the foreground and background of traffic targets under special weather conditions;3)Only a single threshold was selected to distinguish the positive and negative samples,and the detection effect is not ideal.In this thesis,the Faster RCNN algorithm with high accuracy is selected,and we adopt the method of gradually increasing the network module to improve the Faster RCNN algorithm,so that it can be better applied to the target detection in the traffic scene.The main research contents are as follows:(1)Replace the backbone network with a small Res Net50,deepen the network level,extract high-level semantic features,and avoid the problems of information transmission loss and gradient disappearance in the feature extraction process.(2)Aiming at the problems of background noise interference and the difficulty in distinguishing the target background,a deformable convolutional network is adopted to replace the single regular square convolution,so that the convolution kernel can adapt to the morphological and scale changes of different targets,and extract more relevant features with the real target as much as possible.(3)Aiming at the problem that the spatial position of the target in the scene is complex and the feature is difficult to extract,the spatial attention mechanism is introduced to correct the spatial position of the detected target and normalize the arrangement of the target features,so as to improve the detection accuracy.(4)The feature pyramid network is introduced and further improved to PANet to enhance the multi-scale fusion information of FPN.Finally,the better feature integration method BFP is adopted,which makes the comprehensive feature have balance information from each resolution,and relies on the integrated balanced semantic features to enhance the original features.Feature pyramid can effectively solve the problem of small and dense target proportion in image.(5)Cascade RCNN network is introduced to cascade multiple RCNN networks based on different Io U thresholds to continuously optimize the detection results,which can minimize over fitting and improve the detection accuracy.On the COCO dataset,the improved Faster RCNN object detection algorithm proposed in this paper is compared with several main object detection algorithms.At the same time,it is verified on KITTI and Pascal voc2007 + 2012 datasets.Several experimental results show that the improved algorithms can effectively improve the accuracy,and the detection effect is significantly improved compared with the previous object detection algorithm,which has practical research significance.