Research On Key Technologies Of Pedestrian Detection

With the development of intelligent transportation and self-driving technologies,pedestrian detection has attracted wide attention.Pedestrian detection aims at locating the pedestrian positions in the images or videos.Its accuracy is affected by many factors,such as occlusion,multi-scale,shooting angle and illumination conditions.In recent years,convolution neural network has made great breakthroughs in general object detection tasks.How to solve the problems of pedestrian detection based on convolution neural network and effectively improve the accuracy and robustness of pedestrian detection algorithm has important research significance.In this paper,the problem of occlusion is taken as a breakthrough to improve the performance of pedestrian detection.By making full use of the learnable characteristics of convolutional neural network,the effective way to improve the performance of pedestrian detection is explored from the perspective of multi-stream feature fusion and multi-task learning.In this paper,we propose a multi-stream region proposal network to enhance the ability of pedestrian detection.First,multiple visible region guided networks are proposed to obtain the disparate features based on diverse visible region patterns.Then,fusion network is introduced to integrate the feature maps from multiple visible region patterns.Finally,a specially tailored region proposal network is applied to generate the proposal regions using the fused region features.In addition,boosted forest is used to classify the proposal regions.The proposed method is evaluated on Caltech database and achieves comparable performance with the state-of-the-art methods.To optimize the above method,this paper further explores the role of multi-task learning in pedestrian detection.A multi-task learning framework based on region proposal network is constructed by inheriting the multiple visible patterns.In the model training,different visible patterns of samples are passed to the heads of region proposal network which are based on the same backbone network structure to form multiple training tasks,and the proportion of visible patterns is used as the weight of loss function for each task.Since there are inherent correlations between different occlusion types of regional proposal tasks,different tasks can promote each other.Experiments show that our method can improve the performance of the detection algorithm,and the detection ability of occluded samples is more prominent than single-task learning.