Leveraging Human Segmentation Guided Frames in Videos for Activity Recognition
Shaurya Gupta, Dinesh Kumar Vishwakarma, Nitin K. Puri · 2022 6th International Conference on Computing Methodologies and Communication (ICCMC) · 2022
Human Activity Recognition (HAR) has been an essential task in Machine Learning for identifying the concurrent activity through some form of sequential data like in video frames or activity sensors. This research provides a framework, working on video data, that aids the use of an independent deep learning pipeline working on human image segmentation for pre-processing each passing video frame sent into a Convolutional-LSTM HAR pipeline for determining the ongoing activity. The applied pre-processing makes the foreground pixel intensities lower than that of the human subject making the subject appear in better focus. The applied novel pre-processing technique enhances the results compared to ones using standard RGB frames sent to the conventional HAR pipeline. The work is then validated over two publicly available datasets, namely: KARD and MSR-Daily Activity. The results show some superior accuracy compared to other state-of-the-art works over the validated datasets with accuracy as—99.07% (KARD) and 98.75% (MSR-Daily Activity).