| With the increasing demand for public safety,the issue of bus safety has attracted more and more attention.For the video surveillance system of buses,it is very important to ensure real-time detection and understanding of abnormal behavior of moving targets.Current bus video surveillance systems mainly focus on the detection and tracking of moving targets in the scene,but there are still shortcomings in further identifying and understanding the targets.If the system can detect and understand sudden events and other abnormal behaviors in real time,it can respond more quickly and take appropriate action.Based on this,this paper combines image processing,background modeling,clustering algorithms,and posture estimation techniques to study technologies and algorithms that can detect and understand abnormal behaviors of moving targets in real time,in order to improve the effectiveness of the bus video surveillance system.The main work of this paper is as follows:(1)Utilizing the distinctive attributes of the bus interior environment,a smart video surveillance system is devised specifically for buses.The system’s central processing unit for video analysis is built using the NVIDIA Jetson Xavier NX module,which is integrated into the bus platform.(2)For motion feature extraction,a novel human body detection algorithm is proposed,which combines object detection and pose estimation.This algorithm is tailored for bus scenes,effectively addressing challenges like severe occlusion and difficulty in obtaining complete human features.The proposed approach involves two main steps.Firstly,the object detection algorithm is utilized to identify the human target’s bounding box coordinates.Following this,the pose estimation algorithm is employed to identify the key points of the human body.Subsequently,a fusion strategy is developed based on spatial alignment,merging the outcomes of both algorithms to extract precise motion features from the subject.For validation purposes,a tailored dataset is compiled,encompassing datasets for human subjects and their keypoints.This dataset is then used to train and validate the algorithm’s performance.Notably,the experimental findings from human detection highlight the advantages of combining the human detection algorithm,providing a broader scope of global and local features compared to utilizing a single algorithm.This fusion approach enhances the system’s proficiency in accurately recognizing and representing human motion features,particularly in complex bus scenarios.(3)This paper introduces an algorithm for detecting abnormal behaviors in a bus compartment scene using motion features.The proposed algorithm encompasses the establishment of fall detection criteria and the formulation of a behavior discrimination network.The fall discrimination criteria are determined based on the aspect ratio and spatial position of the human body detection frame.Experimental findings corroborate that the proposed approach outperforms single discrimination methods in terms of accuracy.The integration of the discrimination network and criteria,coupled with the result selection strategy,provides an effective means of detecting abnormal behaviors based on motion features,especially in the complex environment of a bus compartment scene. |
