Study On Small-scale Pedestrian Detection Algorithm For Railway Scene And System Design

Railway transportation is China's main mode of transportation and occupies an important position in the field of transportation.Personnel interference in key railway areas is an important factor that affects the safety of railway transportation.Therefore,the study of pedestrian detection algorithms in railway scenes is of great significance to railway transportation safety.Small-scale pedestrian detection is the focus and difficulty of the pedestrian detection algorithm.The main work of this thesis is as follows:(1)Aiming at the problems of high resources consumption and low detection efficiency in the traditional railway system,a small-scale pedestrian detection model based on convolutional neural network(CNN)is constructed.According to the analysis of the general pedestrian detection algorithm and the characteristics of the actual railway scenes,based on the Faster R-CNN(Faster Region CNN)model,pedestrian detection in the railway scene is realized.For the problems in the data set,the small-scale pedestrian images of the railway scene are expanded.Finally,this study designs the structure of each part of the CNN,determine the loss function and training strategy,and complete the construction of Faster R-CNN model.(2)According to the problem of weak feature expression ability and high miss detection rate for small-scale pedestrian,the detection algorithm of small-scale pedestrian based on Faster R-CNN is optimized.First of all,in terms of the problem about low matching between the detection target and the anchor,based on the improved algorithm of K-means and Mean shift,an adaptive candidate network(ACN)is given,which improves the efficiency of generating candidate regions.Secondly,for the reason why small-scale pedestrians are difficult to detect,this study combine shallow detailed information and deep semantic information,and implement pedestrian detection based on the multi-level feature fusion,which increases the target's feature expression ability.Finally,for the problem of less tag data in the data set,this study implement mixed training of unlabeled data and labeled data based on semi-supervised learning(SSL),which helps to obtain more pedestrian information.Experimental results show that the algorithm can significantly reduce the missed detection rate of small-scale pedestrians in railway scenes and achieve good detection results.(3)Based on the pedestrian detection algorithm,a pedestrian detection system in railway scenes is designed.It combines data collection,model training,model reasoning through the control interface to form a pedestrian detection system.The pedestriandetection function of the system is verified by test of pedestrian image acquisition and processing.The system can be used for pedestrian detection in railway scenes.