Pedestrian Detection Under Occlusion Based On Part Detector

With the advent of the era of artificial intelligence,all walks of life have set off a wave of artificial intelligence.Pedestrian Detection is an important research direction in the field of computer vision.As a key technology in autonomous driving,intelligent monitoring and human-computer interaction systems,pedestrian detection has been the focus of academia and industry.At present,most pedestrian detection algorithms can have a good detection effect in the case of a single scene and fewer pedestrians,but as the degree of occlusion increases,the pedestrian detection effect drops sharply.Aiming at this problem,this dissertation takes the pedestrians under partial occlusion as the main research object,gives a pedestrian detection method based on human body parts,and carries out related experimental research.Based on the network structure of candidate region detection and the distribution of human visible regions under occlusion,this dissertation presents a pedestrian detection method based on human body parts.Aiming at the particularity of pedestrian shape and posture,this dissertation designs a detection network with higher adaptability to human body structure based on the general target detection network.By analyzing the distribution of visible regions of pedestrians in the dataset,this dissertation proposes a method for different levels of pedestrians according to the degree of occlusion.The smaller the visible area,the lower the level,and the high-level visible area is composed of multiple low-level visible areas.According to the results of the division of human body parts,a pedestrian-detection network with multiple receptive fields was designed.The multi-scale convolution operation was used to extract the feature information of different levels.In the estimation of visibility for each part,this dissertation proposes a method of “multilayer detection and mutual verification” for the same part of the area.Then we correct the visibility estimation of the part by the back-propagation algorithm of the neural network,which effectively reduces the rate of missing rate of pedestrian detection.Based on the part detection algorithm,to further reduce the influence of occlusion on the detection effect,we optimize the network structure and loss function.In order to obtain more representative features,this dissertation designs two networks of attention mechanisms,so that the model can pay more attention to the region of interest.Aiming at the mutual occlusion of two pedestrians,this dissertation designs a loss function with penalty term,which effectively weakens the influence of the prediction bounding box towards to the non-target bounding box on the detection effect.The model presented in this dissertation evaluates the test results under three test subsets of the Caltech pedestrian dataset,which calculates the MR-FPPI curve values respectively.The missed rate on the Reasonable subset is 12.73%,Partial occlusion subset is 18.89%,and the missed detection rate on the Heavy occlusion subset is 54.68%.Compared with other algorithms and self-contrast experiments,the miss rate of the algorithm is reduced by 1.1% on the Partial occlusion subset and is reduced by 6% on the Heavy occlusion subset,indicating that the network model of this dissertation can better solve the occlusion problem in pedestrian detection.