Research And Implementation On Human Abnormal Behavior Detection Using RealSense

In many public places and security field,the ability to recognize complex human behaviors is particularly important,especially for the detection and early warning of human abnormal behavior.At present,people can know the dangerous events by monitoring multi camera views simultaneously,which is waste of time and inefficient.The rapid development of computer vision makes it possible to automate the process and alert the operator when dangerous situations happen.Such is the case,we mainly take the behavior analysis of holding knife in hand as the main research content and describe the related work detailly.Then we intoduce the RealSense platform architecture,SDK module and the principle of depth image acquisition,which prepares for the follow-up experiments.In order to improve the recall rate of hand-held knife detection method based on RGB image,two improved methods based on MPEG-7 EHD feature extraction algorithm are proposed,which are feature rotation method and additive maximum global edge sum feature method.The feature rotation method adapts different shooting angles by rotating the MPEG-7 EHD feature of the candidate picture.For the long and straight edge feature of knives,the maximum global edge sum feature addition method is proposed for extracting the global edge feature.And this new feature is added after the original MPEG-7 edge histogram descriptor feature to form an81-dimensional feature vector.Experimental results based on public knife detection database and self-made dataset show that the two methods improve the recall rate,accuracy and F-score of hand-held knife detection greatly.The integrated model even shows the highest recall rate,which increasing by 17.56% compared with the basic model.We can see that our algorithms have a better effect on the knife in hand detection problem.Finally,a complete human abnormal behavior detection system is built based on the trained model,which will give alert signals when dangerous event happen in short range from the RealSense SR300 camera.Meanwhile,in order to make full use of the depth image from the RealSense camera,this paper presents a new algorithm for human abnormal behavior detection based on depth image with RealSense R200 camera.In this method,we take a collection of 3D CAD models and render each CAD model from hundreds of viewpoints to obtain synthetic depth maps.For each depth rendering,we extract features from the 3D point cloud and train an Exemplar-SVM classifer.In this experiment,the robustness of the model is also taken into consideration and several methods are proposed to deal with depth and boundary missing,clutter and occlusion.The experimental data is divided into three classes and the results show that the algorithm has a reliable accuracy for hand-held knife detection under a long distance.The recall rate of knife detection even achieves 75% within 3 meters.