Suspicious Human Activity Detection Using Pose Estimation and LSTM
Rohit Nale, Mahesh D. Sawarbandhe, Naveen Chegogoju, Vishal R. Satpute
Abstract
Recognition of both human abnormal activity and daily activity is achieved distinctly with skeleton inputs where geometrical relations of skeletal joint coordinates are observed. These handcrafted features play a significant role to enhance performance. In this paper, the key contribution is that the datasets are reduced to eliminate the complexity of the model by implementing RNN-based approaches. 11,376 skeleton features are reduced to 5 geometrical features selected from 40 different subjects and 12 different action classes of daily and health-related actions. Further, we also introduce an LSTM model to classify medical suspicious activities. The set of relevant geometrical attributes are carefully chosen to train the LSTM model. Experimental work is performed on the multilayer LSTM framework to have a good impact on performance as human activity has a greater diversity. The relations between geometrical different datasets such as selected lines and distances between relevant coordinates of joints are evaluated using NTUD 60 dataset. Also, it is demonstrated that less training data is required for classification between normal and medical activity. The Joint-Joint Distance features show 89.07% accuracy on cross-view evaluation and 78.31% accuracy on cross-subject evaluation.