Research On Pedestrian Detection Based On Convolutional Neural Networks

As one of the fundamental and challenging problems in computer vision,pedestrian detection has attracted much attention during the past decades.In mathematics,as an application for non-convex optimization,the pedestrian detection algorithm based on deep neural network model has made remarkable breakthroughs.Despite this,when facing complex real-life scenarios,there is still no such a model that can both satisfy the accuracy and speed requirements of detection.In terms of this research bottleneck,this thesis focuses on two detection challenges:(1)multi-scale pedestrians detection;(2)severely occluded pedestrians detection.We improve the detection accuracy of the algorithm in the above two types of complex scenes under the premise of ensuring high computational speed.The pedestrian detection methods based on deep learning can be classfied into two-stage and one-stage methods.Since the two-stage methods need to generate proposal regions before feature prediction,the detection speed of this method is generally limited.Though the one-stage method is superior in speed,the information of the feature maps extracted from this method is insufficient,which leads to a suboptimal performance on small-scale pedestrian detection tasks that require the underlying information.Besides,the traditional pedestrian detection methods only predict the feature map once,which results in insufficient positioning ability on small-scale pedestrians,especially in the case where pedestrians are severely occluded.To tackle these issues,we present a one-stage method with multi-level semantic fusion and multi-stage predictors.The method proposed in this article has improved in two aspects compared with the traditional one-stage methods:(1)Fuse multi-level semantic information to enrich the expression of features.Specifically,we first put the original image into the feature extraction network with semantic fusion auxiliary network.Then,the semantic information of the upper-layer features is integrated into the lower-layer features from top to bottom.Finally,we output the features that incorporate both the local and global semantic information.Our method is evaluated on Caltech pedestrian detection benchmark and the experimental results show that our auxiliary network can bring a lower detection miss rate.(2)Multi-stage prediction to strengthen the positioning ability of detectors.First,our model receives feature maps of different scales from the feature extraction network,and then transfers them into multi-stage predictors.The predictors at each stage are optimized by the prediction results from the previous stage,thus our model can gradually push its predicting coordinates to the pedestrian's true position stage by stage.The results show that the method proposed in this thesis can improve the detection accuracy with a small loss of speed.Further analyses show that our model outperforms the traditional one-stage methods on all Caltech test subsets,including that with different scales and different occlusion conditions.Furthermore,the detection performance of our model on the CityPersons validation dataset is also superior to the traditional method.Finally,when compared with the state-of-the-art two-stage method,the model presented in this thesis can achieve competitive accuracy while maintaining its advantage of speed.