Research On The Detection Of Abnormal Behavior Of Passengers In Elevator Car Based On AlphaPose

As a necessary facility for high-rise buildings,elevators make our lives more convenient and faster,but it also brings many safety issues.Because the elevator car has the characteristics of closed environment,small space and opacity,it is easy to become a place for criminals to commit violence.In addition,if a passenger faints or attacks the door when he is alone on the elevator,it cannot be discovered in time by the outside world,which will pose a threat to the life and property safety of passengers.Although elevator cars are generally equipped with monitoring equipment,the massive monitoring data and the sparsity of abnormal behaviors make manual monitoring unable to efficiently and timely detect abnormal behaviors.Aiming at the three kinds of abnormal behaviors that are most likely to occur in the specific scene of the elevator car: single person attacking the door,single person falling to the ground and violent conflict between multiple persons,a passenger abnormal behavior detection method based on AlphaPose is proposed in this dissertation.The method can filter out background interference,focus on the passenger’s posture,and use different algorithms for single-person scenes and multiperson scenes according to the number of people on the elevator to achieve real-time accurate detection of abnormal behaviors.The main research contents of the dissertation are as follows:(1)In the target extraction stage,the key area comparison method is used to judge whether the elevator car is carrying passengers.This method is based on the improvement of the background difference method,which can better adapt to different elevator car environments and improve the accuracy of judgment.When passengers are detected in the elevator car,the YOLOv3 algorithm is used to detect and count the number of people in the elevator.At the same time,the passenger target area is extracted and sent to AlphaPose for human body posture estimation.The estimated results will be used as foundation for subsequent abnormal behavior analysis.(2)According to the characteristics of the single-person abnormal behavior in elevator car,an abnormal behavior detection algorithm based on support vector machine is proposed.The angle,proportion,relative position and other characteristic data of posture are calculated by the coordinates of key points obtained from human posture estimation,and the SVM posture classification model is constructed to complete the classification of three kinds of posture of single person standing,attacking the door and falling to the ground,so as to judge whether the abnormal behavior of single person occurs in the car.On the data set constructed in this dissertation,the abnormal behavior detection rate of the detection algorithm reaches 96.67%.(3)According to the characteristics of multi-person abnormal behavior in the elevator car,an abnormal behavior detection algorithm based on the temporal and spatial characteristics of human posture is proposed.The pose map is reconstructed from the results of human pose estimation,and the spatial interest points of the pose map are extracted by SIFT algorithm.The pyramid L-K optical flow algorithm is used to track the optical flow information changes of the interest points in the time dimension,and then determine whether violence occurs in the elevator car.On the elevator scene data set constructed in this dissertation,the recognition accuracy of the detection algorithm for abnormal behavior is 94.49%;On the public data set of non-elevator scene,the accuracy of abnormal behavior recognition is 90.80%.The method proposed in this dissertation can meet the real-time requirements when detecting the abnormal behavior of passengers in the elevator car,and can provide warning information in time when abnormal behaviors are detected,so as to ensure the safety of passengers.