Occlusion Pedestrian Detection Combining Semantics With Attention Mechanisms

Pedestrian detection is one of the important research topics in the field of computer vision.It's a task that classify pedestrians and backgrounds in video sequences or images and give the specific locations of pedestrians.In recent years,the object detection framework based on deep learning has achieved great success.Pedestrian detection is one of the popular research topics in object detection tasks,and its performance has also made great breakthroughs.However,the general object detection algorithm and pedestrian detection algorithm detect different targets,so the challenges of the two tasks are also different.In complex situations,especially when pedestrians are occluded,background interference,small size,speed of detection and so on.Pedestrian detection algorithms based on general object detection frameworks have a huge progress for improvement.In view of the above problems,this paper conducts research on semantic segmentation,feature fusion,attention mechanism and so on which based on the Faster Region Convolutional Neural Network for object detection framework.This algorithm makes a research on semantic segmentation,feature fusion,attention mechanism.The main work completed in this paper is as follows:(1)We propose a pedestrian detection algorithm combining semantics with multilevel features fusion.Firstly,we extract pedestrian features at different scales by fusing multiple convolutional layer features of the backbone network,which helps to generate robust features with rich details.Secondly,we add a semantic segmentation branch on the fusion layer,and then use the semantic features to connect with the corresponding convolutional layer as the prior information for pedestrian object position to enhance the discrimination between the pedestrian and the background.Finally,a secondary pedestrian detection module is constructed on the preliminary regression of the region proposal network to further exclude false objects.This algorithm is suitable for the detection of pedestrian in intelligent traffic scenes,especially is more robust to occluded pedestrian.This algorithm basically does not increase the computational complexity.(2)We propose a pedestrian detection algorithm combing attention mechanisms with deep features fusion.It has been researched that the different channel features of the convolutional layer usually correspond to different parts of the human body.In this paper,we make full use of these channel information to distinguish the occluded pedestrian from the occluded objects.Firstly,we add attention mechanisms and semantic segmentation branches to multiple convolutional layers.The attention mechanism can automatically selects the semantics and useful information of pedestrians between convolutional layers.The semantic segmentation can effectively assist in detecting and distinguishing pedestrian and background.Secondly,we add a competitive attention mechanism to the final output layer of the backbone network.It is an effectively method to alleviate the problem of redundancy which caused by the increase in the depth of the training network.Finally,the fusion of deep convolution features constitute a comprehensive pedestrian information input region proposal network.By mapping features of different proportions on the convolution layer,the feature information of pedestrians of different sizes is effectively enriched.The algorithm has achieved better performance in the pedestrian detection with different occlusions,especially for the heavy occlusion pedestrian.(3)Our algorithms is validated on Caltech and City Persons datasets and compared with the current mainstream pedestrian detection algorithms.In complex scenarios,the algorithm in this paper improves the performance of occlusion and small size pedestrian detection on the premise of real-time detection speed.