Occlusion Pedestrian Detection Based On Deep Learning Algorithm

With the rapid development of artificial intelligence technology,machine vision has applied into many aspects of our lives,which make people's live dramatic changes.Pedestrian detection,As one of basic identification technologies,has a wide range of application scenarios,such as security monitoring,autonomous driving,and new retail and etc.In this paper,we redesigned the network structure for the specific goal of blocking pedestrians based on the SSD target detection algorithm which greatly improved its detection performance.This article mainly includes the following parts:We firstly make a simple summary of the research difficulties and current situation of pedestrian detection,meanwhile,we comb the traditional machine learning-based pedestrian detection method,while the related knowledge were based on deep learning target detection algorithm.Among those two algorithms,HOG+SVM is the most classic algorithm in the traditional method.And the target detection algorithm based on deep learning mainly includes three categories: RCNN series,SSD and YOLO series.Taking into account the speed and performance of the detection,we trained a pedestrian detection system based on the SSD target detection framework with the self-built occlusion pedestrian dataset.Meanwhile,under the test set and re-labeled INRIA test set,we compared the trained SSD model with the pedestrian detection model,which the latter is based on HOG+SVM and built into OpenCV.The results show that the detection effect of the SSD model is significantly better than the traditional HOG+SVM-based pedestrian detection system.The features learned by the deep convolutional neural network are more robust.Finally,we modified the network structure of the SSD model for the occlusion pedestrian detection.(1)In the pre-network of the SSD model,the SE-Inception structure is added to enable it to extract features more efficiently;(2)Redesign the priori framework in the network to make model easier to match the shape of the pedestrians and adopt a dense sampling strategy for small-sized pedestrians;(3)For the situation where pedestrians are prone to environmental occlusion and mutual occlusion,the data of occlusion pedestrians is added to the training set of the model,at the same time,we use Repulsion Loss to strengthen the ability to detect pedestrian of the model.The final results show that the improved SSD model has a significant improvement in detection performance while the detection time only has a small increasement.