Development Of Prison Sentinels' Pose Fast Recognition Algorithm Based On Keypoints Spatio-temporal Map

Prisons are important places to maintain national security,and their requirements for safety are higher than other places.Sentinel is responsible for guarding and defending,and needs to respond to various abnormal situations quickly.The status of the sentinel is a key factor to ensure the safety of prisons and needs real-time monitoring.The traditional video surveillance systems of prisons rely on manual monitoring the duty status,which consume resources and have low monitoring efficiency.Therefore,developing the fast monitoring algorithm of detecting duty status automaticly has a strong engineering application value.This thesis develops a fast recognition algorithm for the sentinel posture of the duty room based on the keypoints spatio-temporal map,and automatically monitors the sentinel's duty status.The algorithm adopts the principle of fast and lightweight,and is divided into three modules:human object detection,human keypoints detection and action recognition.The human object detection module is based on the single-step detection framework,and uses the shallow full convolution neural network for feature extraction to ensure detection accuracy while reducing model complexity;analyzing and designing the data normalized layer,loss function and other parts in detail;and data enhancement,multi-scale training and other methods are used in the training process to improve detection accuracy.The human keypoints detection module uses a multi-model aggregation method to process some unlabeled samples,alleviating the problem of insufficient keypoints data in the self-built data set;using knowledge distillation technology to compress the complex multi-layer stacked hourglass networks,significantly reducing the amount of model parameters and calculation.The action classification recognition module collects multi-frame human keypoints data to construct a spatio-temporal map for classification and recognition,and fully learns the spatio-temporal and co-occurrence characteristics of keypoints data through two-way feature extraction and channel conversion and uses the global average pooling method to integration features.In this thesis,each module of the algorithm is tested separately and also integrated for overall testing.For different duty scenes,the algorithm can effectively recognize abnormal actions such as putting feet on desk and putting face on desk,and realize automatic alarm.The accuracy rate of classification can reach 99%.On the GTX1080Ti,the algorithm costs about 7.1ms to obtain the keypoints from the input image,and costs about 1.5ms to realize the action recognition by the keypoints spatio-temporal map,which meets the requirements of fast action recognition,and is practically applied in the video surveillance system of prisons.