Research On Tunnel Pedestrian Target Detection Algorithm Based On Convolutional Neural Network

The rapid and accurate detection of pedestrian targets in tunnels plays an important role in ensuring tunnel traffic safety.Compared with traditional machine learning detection algorithms,convolutional neural networks can automatically learn rich features and have strong feature extraction capabilities.Pedestrian detection algorithm based on convolutional neural network has obvious advantages in speed and accuracy.In the tunnel environment,the ambient light is insufficient,the surveillance video image is blurred,and the noise is high.The pedestrian target is small in the tunnel monitoring video,the pixel is low.It is difficult to exert the advantage of strong feature extraction ability of convolutional neural network.Therefore,it is of great theoretical and practical significance to study tunnel pedestrian target detection based on improved convolutional neural networks.In recent years,the convolutional neural network represented by the R-CNN series has achieved good results in target detection.This paper is based on the Fast R-CNN and Faster R-CNN networks.For the problem that the time of extracting the candidate regions of Fast R-CNN is too long and Faster R-CNN has poor image feature extraction in tunnel environment,a new foreground extraction method and a target detection network which cascade super-resolution network with convolutional neural network is proposed.In order to improve the detection accuracy,the RPN network and NMS algorithm in Faster R-CNN are also improved.The main work and contributions of this paper are as follows:(1)Firstly,this paper introduces the Fast R-CNN detection network and analyzes the Selective Search candidate region extraction algorithm used in Fast R-CNN.Aiming at the problem that it is too long to extract candidate regions,a foreground extraction method combining the faster background difference and the interframe difference method is proposed to reduce the detection time.Super-resolution reconstruction enhances the detection accuracy by supplementing the high-frequency information of the image and increasing the image detail information,so that the Faster R-CNN generates image features with richer semantic information.(2)Visualization analysis of image features by deconvolution operation reveals that the convolutional neural network has insufficient feature expression ability for low-resolution pedestrian target extraction in tunnel environment.Aiming at this problem,a new SR-CNN pedestrian target detection network which cascaded super-resolution network and Faster R-CNN is proposed.The method in this paper supplements the high-frequency information of the image based on super-resolution reconstruction,and increases the detail information of the image,so that Faster R-CNN generates feature maps with richer semantic information,thus improving the detection accuracy.(3)In Faster R-CNN,when the RPN network extracts the candidate region,the Anchor size of candidate frame is manually designed,and the pedestrian prior dimension information is not utilized,resulting in the candidate window being extracted is not accurate enough.Aiming at this problem,this paper uses K-Means algorithm to cluster and measure the size and proportion of the actual frame of pedestrians to generate higher quality candidate windows,and improve the accuracy of prediction box regression,thus improving the detection accuracy.(4)When dealing with overlapping windows in Faster R-CNN algorithm,the greedy algorithm based on fixed threshold will delete the overlapping window exceeding the threshold,which will easily lead to missed detection of overlapping targets.In response to this problem,an improved NMS algorithm is used in this paper,setting a fractional decay function to reduce the score of overlapping windows instead of directly zeroing to avoid direct deletion of overlapping targets.This method can reduce the missed detection of overlapping targets effectively.Based on the above improvements,a New-CNN pedestrian detection model based on improved convolutional neural network is formed.The paper uses the dataset which maked by the monitor video on Chongqing Highway Tunnel,trains and tests the improved models,and conducts a number of comparative experiments.The experimental results show that compared with the original Fast R-CNN and Faster R-CNN algorithms,the proposed method improves the detection speed and detection accuracy,and achieves better detection results in the actual environment.