Human Abnormal Behavior Detection Based On Deep Learning

In the aspect of human abnormal behavior detection,the research on combination of deep learning is the current frontier research topic.Deep learning is affecting security companies.In various security monitoring scenarios,intelligent monitoring has gradually become a development trend.As an important place for scientific research units,laboratories naturally have a variety of complex and diverse safety monitoring scenarios.At the same time,researchers in the laboratory have become the main targets for safety monitoring.Therefore,in this project,a complex laboratory is used as a testing scenario to apply deep learning to human abnormal behavior detection in monitoring,and to detect abnormal behaviors of laboratory staff so that abnormal human behaviors can be detected in real time.So as to establish a foundation for subsequent security measures such as early warning.In addition,it is also important to improve the level of laboratory intelligence and level of security.The main contents of this paper are as follows:(1)According to experimental requirements,in order to improve the accuracy of human abnormal behavior detection in specific scenes,personal experimental data set of human abnormal behavior is built.This topic is researched based on the data in security monitoring.Due to the adoption of deep learning methods,data sets need to be established.At the same time,due to specific scenarios,existing data sets cannot meet the experimental requirements.The experiment used monitoring data in an onsite internet studio to detect defined abnormal human behaviors.Through a series of operations such as frame cut of video,image selection,and label calibration,to build a training and test data set of human abnormal behaviors.(2)For the problem that the human body's abnormal behavior is difficult to detect in complex laboratory environments,the abnormal behavior is classified directly to establish a dataset in the experiment,and then train YOLO-based network structures to obtain detection models(YOLO-Abnormal Behavior Detection,abbreviated to YOLO-ABD).The main steps of the establishment of YOLO-ABD model are as follows: Automatic feature extraction and classification is performed by inputting complex anomalous behaviors into the deep neural network YOLO,and the target extraction step is given to the neural network,to the extent that the target classification can be put into a network at the same time.With the use of deep neural feature extraction and high-precision detection of classification features,it is possible to accurately detect defined abnormal behaviors and achieve end-to-end abnormal behavior detection that from input data to output detection results.Through the verification for acceleration of the detection process by the GPU,real-time detection requirements of the surveillance video can be satisfied.In this paper,according to the definition of the abnormal behavior of the monitoring scene,an experimental data set of abnormal human behavior is established.The target detection network model YOLO is applied to detect abnormal human behavior.The YOLO-ABD network model which obtained through training network can detect anomalous behaviors well.In the experiment,the recall rate has reached more than 90%,far from the supermarket Face ++ platform human detection 75.7% recall rate,and the average precision rate was over 95%,in the meantime,it is verified with a simple model.In the case of experimental GPU acceleration,the detection speed of the video stream can reach about 30 FPS,which satisfies the requirements of real-time detection requirements.