The Design And Implementation Of Escalator Intelligent Video Surveillance System Based On TX2

The escalator can cause some safety problems while it is convenient for passengers to travel.It needs to be monitored to avoid the occurrence of safety accidents.Intelligent video surveillance has obvious advantages over manual surveillance,such as stable surveillance,fast response and low cost.It has been paid more and more attention in the field of safety prevention in public places.The purpose of this article is to design and implement an intelligent video surveillance system for escalators based on TX2 embedded platform.The escalator scenario was monitored by three cameras with different shooting directions.Passengers on escalators and floor boards were detected and tracked to realize passenger detection,passenger flow statistics and crowd abnormal behavior detection.Passenger skeleton was extracted to realize passenger abnormal behavior detection and abnormal behavior recognition.At the same time,passenger congestion detection and retrograde detection were carried out on the floor boards,and foreign body extension detection was carried out on the handrails.The mian work of this article can be divided into the following four aspects:1)Passenger detection:Firstly,the common target detection methods of foreground extraction,machine learning and deep learning were used to detect the passenger heads on floor boards and passenger faces on escalators to analyze the advantage and limitation of various methods,and finally we adopted the LSVM classifier with DPM features as passenger detection method and selected the best parameters of the classifier through comparative experiments.At the same time,according to the characteristics of foreground extraction method,the foreground of the handrail was extracted by gaussian mixed model to complete the foreign body extension detection on the handrails.2)Passenger tracking:Firstly,the advantages and limitations of common target tracking methods in the escalator scenario were analyzed.Then,the improved KCF method was used to track passengers.Based on the improved nearest neighbor target matching strategy,the tracking target was corrected by the target detection results at intervals,and the detection confidence model was established for the target to measure the possibility of being a passenger.Finally,we calculated the motion characteristics of the tracking target and monitored the passenger behavior based on the motion characteristics ti complete passenger detection,passenger flow statistics and crowd abnormal behavior detection on the escalators,as well as passenger congestion detection and retrograde detection on the floor boards.3)Abnormal behavior detection and recognition of passengers on the escalators:Firstly,the human skeleton feature was selected to describe the behavior of the passengers on the escalators,and then the two-dimensional skeleton of passengers were extracted by OpenPose deep learning network.Then,the direction cosines of the trunk of the skeleton were calculated,and the human posture eigenvectors were combined to determine abnormal skeleton based on template matching,completing the abnormal behavior detection of passengers on the escalator.Finally,the abnormal behavior fragments were segmented from the video,abnormal skeleton was combined into abnormal skeleton sequence,which was matched with various abnormal behavior templates by dynamic time regulation,and abnormal behavior recognition was completed based on K-nearest neighbor prediction behavior category.4)System implementation and testing based on TX2:Firstly,the software implementation of the system was introduced,and then the transplantation of TX2 was explained,including fixing the number of multi-camera equipment,packaging and publishing applications and optimizing the stability of the program.Finally,the system effect was tested.The test shows that the system can accurately and real-time complete the required monitoring tasks and has engineering application value.