Research On Object Detection Algorithm Of Traffic Scene Based On Deep Learning

The field of autonomous driving is receiving more and more attention from industry and academia,and in the field of automatic driving,object detection is one of the core issues.This paper mainly studies the problem of small object detection and occlusion object detection in the object detection field under traffic scenes,so that the vehicle can accurately detect the traffic identifier and vehicle on the road during the driving process.For the difficult problems in the detection of two object,small object and occlusion object,this paper uses two different processing methods to solve.For small object detection,such as traffic sign,this paper proposes to use the attention mechanism to narrow down the areas that need attention,so as to solve the problem that information loss occurs during the forward propagation of convolutional neural networks..In this paper,the feature extraction network is improved,and after the feature extraction,the decoder is added for information fusion,and the detection result is finally improved.In this paper,TT100 K and GTSDB datasets are selected for experiments.In the TT100 K dataset,the accuracy of the proposed algorithm reaches 85.2%,which is 2.1% higher than that of Faster RCNN.Experiments show that the attention mechanism can effectively improve the detection of small object.For the vehicle detection problem,mainly due to the occlusion between the vehicles,the information is lost,and the network cannot accurately determine the object position.Therefore,this paper proposes to use the context coding module to extract the surrounding information of the object,which is helpful for object detection.In this paper,experiments are carried out on KITTI and Udacity datasets.Experiments show that the proposed algorithm achieves 89.5% accuracy on the KITTI dataset,which is 1.6% higher than Faster RCNN,and 0.3 higher than MS-CNN.%.The experimental results show that the context information can effectively improve the detection accuracy of occlusion object in complex scenes.