Research On Occluded Pedestrian Detection With Relation Mining

Pedestrian detection is one of the core technologies of intelligent perception.It is an indispensable technology in a variety of fields,such as video surveillance,intelligent driving,and human-computer interaction.In recent years,with the rapid development of deep learning,the performance of pedestrian detection methods based on deep learning improves significantly.However,in complex scenes,the performance of occluded pedestrian detection can not meet the requirements of practical applications.Therefore,pedestrian detection needs innovation and breakthrough.This thesis improves the pedestrian detection performance by mining various relations.Specifically,this thesis conducts research on mining the relations between visible regions and full-body regions,the relations between different detection boxes,the relations between parts,and the complementary relations between multi-level features.The main contributions of this thesis are as follows.(1)Occlusion results in the information loss of pedestrians,aggravates the intra-class differences of pedestrian features,and leads to the low detection accuracy of occluded pedestrians.To tackle this problem,this thesis proposes a mask-guided feature enhancement model.The proposed model explores the relations between visible regions and full-body regions,thereby highlighting the features of visible regions,and suppressing the occluded regions,so as to improve the support of the visible regions for the existence of the pedestrian and reduce the noise introduced by the occluded regions in the pedestrian features.Experimental results show that the proposed model improves the detection performance of occluded pedestrians.(2)In the crowd scenes,the intra-class occlusions lead to the low detection accuracy of occluded pedestrians.To tackle this problem,this thesis proposes an occluded pedestrian detection method that mines relations between different detection boxes.The proposed method mines the between different detection boxes to obtain the pedestrian count and the detection box similarity,and integrates them into the two-stage pedestrian detection method to improve the detection performance in crowd scenes.Experimental results demonstrate that the proposed method improves the detection performance of occluded pedestrians.(3)The existing part-based pedestrian detection methods lack of mining the intrapart relations and the inter-part relations.To tackle this problem,inspired by human perception mechanism,this thesis proposes an occluded pedestrian detection method based on part co-occurrence relations mining.The proposed method implemented by the graph convolution network enhances the pedestrian features by mining both the inter and intrapart co-occurrence relations,thus improving the detection performance of the occluded pedestrians.Experimental results show the effectiveness of proposed method.(4)The feature pyramid network is the most effective way to tackle the problems of multi-scale pedestrians.But the existing methods usually fuse multi-scale features by simple aggregation manners(such as element-wise addition,concatenation,etc.),which leads to the introduction of irrelevant information in the fusion process.To tackle this problem,this paper proposes a fully connected and complementary network.The proposed method fuses multi-level features adaptively by mining the complementary relations.It can enhance the representation of multi-level features,thus improving the detection performance of multi-scale pedestrians.Experimental results demonstrate that the proposed method obtains accuracy improvements with negligible overheads.