Research On Occluded Pedestrian Detection In Crowd Scenes And System Implementation

Pedestrian detection is an important research direction in the field of computer vision,aiming to locate and identify pedestrian targets in still images or surveillance videos,and is an integral part of the intelligent surveillance field.In recent years,with the rapid development of deep learning,pedestrian detection technology has made great progress,but when it is applied to realistic crowd scenes,the problem of in-class occlusion in crowds and the interference problem of noisy background seriously affect the accuracy of pedestrian detection,resulting in its failure to meet the needs of practical engineering applications.To address such problems,this paper carries out research on the topic of obscured pedestrian detection in crowd scenes to solve the problem of inability to accurately detect obscured pedestrians,further enhance the intelligent and autonomous data processing and analysis capabilities of detection and monitoring devices.Therefore,this paper designs a spatial correction module based on External Attention,and constructs a pedestrian detection method based on dual-level SCenterNet mask feature enhancement.To complete the accurate detection of pedestrian targets,a pedestrian detection monitoring system is designed and developed.The specific research is divided into the following three aspects:(1)SCenterNet detection model incorporating Improved External Attention ModelFirstly,to address the problem of high computational cost and complex structural redundancy of External Attention,a lightweight spatial correction module is designed.It can avoid the influence of channel downscaling operation on the learning effect of Attention Mechanism and enhance the ability to resist noise interference.Secondly,to address the problem of insufficient encoding of local feature information in crowd scenes,this paper designs a two-stage prediction of SCenterNet.In the first stage,it performs prediction regression through the Center Attentive Module and returns the rough box position of the corner points in combination with the Attention Transitive Module.In the second stage,the Aggregation Attentive Module is used to further refine the target prediction box to obtain a more accurate one for pedestrian targets.The feature regression with multiple key points encodes more local information of the target in complex scenes.Finally,SCenterNet detection model incorporating Improved External Attention Model is constructed to improve the ability of the model to encode feature information in crowd scenes.(2)Pedestrian detection method based on dual-level SCenterNet mask feature enhancementA dual-level SCenterNet mask feature enhancement pedestrian detection method is designed for the problem of feature interference and feature information missing among pedestrian targets in crowd scenes.Firstly,an occluded pedestrian detection mask mechanism is established to suppress the feature information of intact pedestrian targets,enhance the local feature information of occluded pedestrians,and attenuate the interference of noisy background and other noises.Secondly,a dual-level detection network is used to detect intact and occluded pedestrian targets in a hierarchical manner,which effectively reduces the interference of similar features among pedestrian targets and improves the detection performance of the model in crowd scenes.Finally,this paper introduces a new post-processing mechanism to replace the traditional non-maximal suppression method,which improves the post-processing capability of the model in crowd scenes by calculating the Manhattan Distance to comprehensively evaluate the confidence of candidate box.Through multiple sets of ablation and comparative experiments,it has been shown that the proposed method in this paper performs better than other detection models in crowd scenes,with strong generalization and robustness.(3)Design and implementation of a pedestrian detection and monitoring systemIn this paper,we design and implement a pedestrian detection and monitoring system with a pedestrian detection model containing dual-level SCenterNet mask feature enhancement as the core.The system can use multiple detection models to achieve accurate pedestrian detection in static images and surveillance videos.Based on the characteristics of crowd scenes,statistical analysis of real-time pedestrian monitoring data is conducted,and a series of characteristic functions in crowd scenes are provided.