Key Technologies For Human Abnormal Behavior Detection In Home Video Surveillance Environment

Since abnormal behaviors indoors(such as falls,fights,etc.)often lead to serious safety accidents,resulting in accidental injuries to people,abnormal behavior detection in home monitoring environments has become one of the hottest research directions.The elderly are the main age group susceptible to accidental injuries caused by abnormal behaviors such as falls.With the aging of China's population,the number of empty nest elderly has increased rapidly,and the supervision of the elderly has become an increasingly prominent social problem.The main research direction of this article is the detection of human fall behavior indoors.The purpose of the research is to achieve effective detection of the fall of humans indoors.Aiming at the problem that the traditional fall algorithm is difficult to effectively extract moving targets due to similar foreground and background during the moving target extraction stage,an improved Visual Background Extractor(Vi Be)algorithm based on multi-information fusion is proposed for moving target extraction,which improves the effectiveness of moving target extraction.After acquiring the moving target,to ensure the detection accuracy and reduce the computational complexity of the fall detection algorithm,a human fall detection framework based on multi-feature fusion combining coarse judgment and fine recognition is proposed,which effectively reduces the computational complexity.The main work and innovations of this article are as follows:(1)For the problem that the moving foreground target is difficult to be effectively extracted when it is similar to the background.This paper proposes an improved Vi Be algorithm that combines top-down information and bottom-up information to enhance foreground detection based on weight membership function.First,the background model is used to capture bottom-up information,and the foreground model is used to capture top-down information.Then,according to the number of matching samples of the pixel and each model and the distance between the pixel and its matching sample,the matching degree between the pixel and the two models is calculated respectively.Finally,a decision framework based on weighted membership using combination of foreground matcing degree and background matching degree is used to determine pixel labels.Experimental data shows that the proposed method is superior to other comparison algorithms used in the subjective and objective evaluation,and that the proposed algorithm is significantly improved in both subjective and objective evaluation compared to the original Vi Be algorithm.Experimental results prove the effectiveness of the proposed method,especially for processing video sequences with similar foreground and background.(2)In order to reduce the computational complexity of the algorithm on the premise of ensuring the detection accuracy,this paper proposes a human-fall detection framework that uses RGB-D images based on multi-feature fusion combining rough judgment and fine recognition.The basic idea of the framework is to first detect the landing behavior of the human during the rough inspection phase,and then determine whether the human landing behavior is actively approaching the ground or falling unexpectedly during the fine inspection phase.The low-complexity coarse detection stage will trigger the high-complexity fine-detection stage only after detecting the human's landing behavior.Therefore,the system will only perform high-complexity calculations when a suspected fall behavior occurs,greatly saving the system's computing resources.At the same time,the fusion of shape features and motion features ensures the detection accuracy of the algorithm.The detection of human grounding behavior in the rough inspection phase is realized by three parts: retentate detection,human recognition,feature extraction and classification.At the fine inspection stage,the final fall judgment is achieved through target tracking,motion feature extraction and threshold classification.The experimental results show that the proposed fall detection framework can effectively detect the fall behavior of the human body while maintaining low complexity.