Research On Pedestrian Detection Algorithm In Dense Scene Based On Deep Learning

Pedestrian detection algorithms have a wide range of applications in areas such as autonomous driving,surveillance,and military.In real life,there are often dense pedestrian scenes where pedestrians are occluded by objects and pedestrians are occluded from each other,which can be categorized as inter-class occlusion problem and intra-class occlusion problem,respectively.Since occlusion interferes with pedestrian features,traditional pedestrian detection techniques usually cannot handle them effectively,so this paper uses a deep learning-based approach to solve these two problems of pedestrian detection in dense scenes.To address the above occlusion problems,this paper investigates algorithms and models for pedestrian detection in dense scenes based on deep learning,and the main contents of this paper are as follows:1. For the common problem of missed detection in inter-class occlusion,a pedestrian detection algorithm that fuses human key points attention feature and the visible part attention feature is proposed.The model constructs two attention modules by introducing human key points and the ground truth bounding box of human visible parts respectively,which suppress the occluded parts in the channel features and spatial features of pedestrian features respectively,so that the MR(Miss Rate)of this model is reduced to 40.78 on the Heavy subset of the City Persons dataset,and the Caltech dataset also obtained a better detection result.2.For the common problem of missed detection in intra-class occlusion,a pedestrian detection algorithm based on sample quality screening of candidate detection bounding box is proposed.The model selects the best candidate detection bounding box matched by each ground truth bounding box by calculating the classification quality and regression quality of each candidate detection bounding box that is a positive sample,while suppressing the confidence of other candidate detection bounding box to reduce the interference of low-quality redundant candidate bounding box on the detection results.In order to optimize the problem of small scale pedestrian targets,a feature context fusion module is proposed to allow the fusion of large scale features with small scale features,which enables the network to perform more accurate localization.The experimental results show that the MR(Miss Rate)of the proposed method on Reasonable in the City Persons dataset is reduced to 13.87 without affecting the inference speed.