Authors : Shadiya Mohammed Raly and Priyantha Kumarawadu
Abstract: The recognition of a burglar caught in CCTV surveillance in real time remains a challenging problem in the domain of action recognition. Existing security systems prioritizes the data acquired from sound sensors, motion sensors, glass breaker sensors over visual sensors to understand the context behind a sequence of action. The proposed system uses the Real-Time Streaming Protocol (RTSP) address of the surveillance camera to acquire the live surveillance images and then uses Open Pose which is a real-time person key point detection library to extract 2D skeletal data which are then fed into a Long-Short Term Memory (LSTM) model, Recurrent Neural Netowrk (RNN) model and Gated Recurrent Unit (GRU) model for classification. The experimental results showed that the performance of LSTM based classifier out performed against RNN and GRU based classifiers under various burglar actions and it was were promising with a training and a validation accuracies of 92.3% and 86.5%, respectively.
Shadiya Mohammed Raly and Priyantha Kumarawadu, 2022. Real-time Burglar Recognition Based on Human Skeletal Data Using Openpose and Long Short Term Memory Network. Asian Journal of Information Technology, 21: 1-5.