Research On Object Detection In Traffic Scene Based On Convolutional Neural Network

Traffic Scene Perception(TSP)is a real-time acquisition of pavement information.It includes three phases: detection of the target of interest,recognition of the target of interest and tracking of the moving target.Since detection is the key to identification and tracking,the target detection algorithm is the key to TSP.In this paper,traffic environment as the detection background,based on Regional convolution neural network(RCNN)visual detection methods,the detection area generation and screening process optimization,in order to improve the existing methods of real-time,effectiveness.Through the detection area optimization,the data of traffic scenes are used to carry out template training to detect and identify the target.On the GPU computer,21 fps traffic target recognition is achieved.In this paper,by analyzing the principle of deep learning,the parameters and functions used in the process of deep learning of convolutional neural networks are studied.Compared with medium and large convolutional neural networks,the detection speed of medium-sized networks reaches 17 fps.In real-time detection,this method is in advantage.This paper analyzes and improves the detection area generation algorithm.In order to improve the real-time performance of the algorithm,by comparing the characteristics of each algorithm generated in the region and the range and characteristics of the target in the traffic environment,this paper analyzes and divides the road region by vanishing point detection and achieves the targeted detection Area generation.In the real-time test experiment,this method improves 7fps on the basis of RPN's 14 fps and the detection frame rate reaches 21 fps,which realizes the high real-time detection of the forward traffic target.In order to improve the effectiveness of the algorithm,this paper analyzes the theory of the frame selection algorithm and the experimental analysis.In view of the existing problems of the principle of NMS detection box screening algorithm,the method of this paper improves the detection effect of the occluded target by means of Gaussian score weighting.The test of the standard dataset verifies that the proposed method achieves a mAP increase of 1.1 and 1.5 on the standard dataset respectively,reaching 73.2% and 70.4% respectively.Faced with the application of vision detection algorithm in autopilot and robot platform,experiments and analyzes were carried out.The visual target detection is taken as an important means of environmental perception to verify the target detection algorithm based on convolutional neural network and the improved method in this paper.In addition,a mobile platform is built to analyze the environmental perception of the mobile platform Application of the feasibility.For the objective of severe occlusion and common method difficult to detect,the proposed method can be tested with confidence of 0.64.For the incomplete target,the method can be tested with the confidence of up to 0.98,and the successful detection with confidence of more than 0.9 of nighttime and shadowed targets also shows that the proposed method is robust to light conditions.