Research And Implementation Of Pedestrian Detection And Attribute Recognition Algorithm Based On Crowd Density Estimation And Attribute Association

With the rapid development of deep learning,intelligent transportation has also become a widely researched and popular field.Pedestrian detection and pedestrian attribute recognition,as key tasks,have also received extensive attention from academia and industry.Both tasks are computer vision tasks with a wide range of application scenarios,and they complement each other in complex and changing real-world scenes.In pedestrian detection,detecting targets that are occluded poses a challenge,and there is still much room for improvement in handling occlusion in existing methods.In the pedestrian attribute recognition task,existing methods often focus more on attribute feature mining and attribute feature localization,neglecting the information between attributes.This project aims to optimize the detection model’s detection ability in occlusion scenes by combining real-world scenarios and utilizing attribute relationships to enhance pedestrian attribute feature expression,thereby improving pedestrian detection and attribute recognition capabilities.In this paper,we propose a pedestrian detection algorithm based on crowd density estimation and a pedestrian attribute recognition network based on attribute correlation.On this basis,we design and implement a pedestrian detection and attribute recognition system.Firstly,in the pedestrian detection task,we introduce a crowd density estimation algorithm to improve the non-maximum suppression algorithm and solve the intra-class occlusion problem.To solve the dilemma faced by the non-maximum suppression algorithm:a higher non-maximum suppression threshold often leads to many false detections,while a lower threshold often leads to the loss of highly overlapping pedestrian targets.We use the crowd density estimation algorithm as a branch of multi-task learning to predict the crowd density in the image and adaptively adjust the non-maximum suppression algorithm’s threshold according to the crowd density.Secondly,in the pedestrian detection task,we introduce a contrastive learning model to solve inter-class occlusion problems.For pedestrian instances that are occluded between classes,the visible part enables the model to detect the existence of pedestrians even in the presence of occlusion.Therefore,the visible part’s features need to be enhanced.Therefore,we plan to introduce a self-supervised model,which strengthens the visual feature expression of the visible part through contrastive learning and weakens the interference of the occluded part on the model,thus enabling the model to perform well even in the presence of inter-class occlusion.Thirdly,in the pedestrian attribute recognition model,we introduce the attention mechanism of Transformer.Existing pedestrian attribute recognition algorithms mostly focus on attribute localization and feature extraction expression,often neglecting the correlation between attributes.This project plans to use the attention mechanism in Transformer to build correlation weights between pedestrian attributes and guide attribute feature expression.For example,if the gender is male,there is a low probability of wearing a skirt,etc.This method can enhance the model’s attribute feature extraction ability and improve the model’s robustness.Fourthly,based on the introduction and research on pedestrian detection and pedestrian attribute recognition algorithms in the previous sections,a pedestrian detection and attribute recognition system is implemented.