Pedestrian Occlusion Detection Based On Full Convolutional Neural Network

In the field of object detection,pedestrian detection is a hot topic with applications such as driverless car auxiliary systems,intelligent monitoring systems and service-oriented intelligent robots.The focus of this thesis is pedestrian occlusion that can be divided into two types of occlusions,including human-to-human self-occlusion and object-to-human occlusion.Human-to-human occlusion is mainly caused by overlapping areas between multiple pedestrians,which cause the prediction frame to be easily offset to other pedestrians around it.The object occlusion is mainly caused by pedestrians being obscured by non-pedestrian objects such as buildings,trees and cars.Resulting in the inability to obtain complete pedestrian information and false or missing inspections.In order to improve the performance of pedestrian occlusion,the thesis mainly deals with the two occlusion situations.Aiming at the occlusion problem of human to human,this thesis designs a human internal occlusion model based on full convolutional neural network,which consists of VGG16,feature pyramid and two full convolution sub-tasks.By introducing the repulsion loss function,the candidate frame is moved away from the adjacent non-target label box.The method can minimize the overlapping area of the prediction frame and other target labeling frames,and avoid the prediction frame from shifting to the area of other labeling frames,thereby effectively reducing the false detection rate of human occlusion.For the occlusion of objects to humans,this thesis proposes a pedestrian occlusion detection model based on semantic attention model,which is optimized based on the occlusion model in the class.It mainly includes a semantic segmentation module and a detection module.The semantic segmentation module uses the visible bounding box of the pedestrian as the label to perform semantic segmentation,in order to obtain the attention map of the pedestrians occluded by the object and the background.The detection module uses the detector to obtain richer pedestrian features from the semantic attention hotspot map.When pedestrians and objects are detected,the detector will focus more on detecting the body part of the pedestrian,allowing more precise positioning of the pedestrian.This paper mainly conducts experiments on the public dataset Citypersons,compares the proposed method with other methods,and the detection effects of the two occlusion models and the attention hotspots obtained by the semantic segmentation network are visualized.The experimental results show that the model is effective for pedestrian occlusion detection and the results are superior to the most advanced methods.