Design And Implementation Of Intelligent Video Surveillance System For Escalator Based On Jetson Xavier

Escalators can be found everywhere in modern cities,bringing convenience to people's lives.With the gradual increase of high-rise buildings and the steady improvement of grades in today's era,it is not only required that elevators be fast and comfortable when they are boarding,but people are increasingly demanding for the safety and reliability of elevators,and escalators are also developing in an intelligent direction.Based on Jetson Xavier embedded equipment,this thesis has developed an intelligent video surveillance system for escalators.The system analyzes the images collected by the cameras taken from three different directions to complete the detection and tracking of passenger heads and large items on the escalator floor,as well as the extraction and tracking of the passenger skeleton on the escalator.On the basis of this,the functions of passenger flow statistics,passenger retrograde detection,passenger congestion detection,passenger behavior recognition,and large item retention detection are realized.The main work of this thesis includes the following four aspects:1)Passenger head and large items detection: After comparing the advantages,disadvantages and limitations of traditional foreground extraction methods,machine learning and deep learning methods,YOLOv3-tiny was finally selected as the passenger head detection method,and KNN background modeling algorithm combined with KD tree algorithm was selected to detect large items.2)Passenger head and large items tracking in overhead scenes: After comparing various tracking methods,Kalman filtering is used to perform motion estimation on the passenger's head,and the Hungarian matching algorithm is used to match the rectangular frame intersection matrix of the tracking result and detection result to match the passenger head.Hungarian matching algorithm is used to match large items in adjacent frames to achieve large item tracking.Based on the tracking results,the behavior of passengers and items on the escalator floor is detected,including passenger flow statistics,passenger retrograde,passenger detention,passenger congestion,and large item detention.3)Passenger abnormal behavior recognition: After comparing the advantages and disadvantages of various skeleton extraction methods,the top-down skeleton extraction method is used.First,YOLOv3 is used for passenger detection on the escalator,and then Mobile Netv2 is used for single-person skeleton extraction.Then use Kalman filtering method to perform motion estimation on the key points of the human body,calculate the distance matrix of the key point skeleton for tracking and extraction,and use the Hungarian matching algorithm to optimize the matrix to achieve passenger tracking on the escalator.Based on the tracking results,a graph convolutional neural network combined with a sliding window is used to identify passenger behavior.4)Implementation and test based on Jetson Xavier: First introduce the structure of the functions of each part of the system software,and then use the CMake tool to generate a makefile,compile each part of the system,and link to generate an executable.Finally,the system is tested for performance.The results show that the system can complete various tasks stably and accurately,and has important use value.