Occlusion resolving inside public crowded scenes based on social deep learning model

Past decade, the field of video analytics has been rapidly developing specially for crowd scenes. The advances in computational resources inspired researchers to build reliable video analytics systems that works real. The main root for any video analytics system is threat activity localization inside video streams. One major issue toward achieving that objective is Occlusion due to crowd intensity. In this paper, a hybrid deep learning model that exploits Convolution Neural Network (CNN) and Social Long Short-Term Memory (LSTM) for real-time video streaming analytics. The experiments were conducted on public available dataset UCY which contains 2 main scenes with 786 persons and 55 actions. The results concluded the superiority of Social LSTM over conventional LSTM and Mean Square Error (MSE) does not exceed 0.25.