Extreme action recognition from real-time video using time-series deep learning model

Goh, Qing Hao (2021) Extreme action recognition from real-time video using time-series deep learning model. Final Year Project, UTAR.

Preview

PDF Download (3410Kb) | Preview

Abstract

The development of an extreme action recognition model to automate police surveillance can improve police deployment speed to crime scenes such as assault, robbery, kidnapping and other offences. However, the existing solution of extreme action recognition is insufficient to be deployed with high confidence. This study proposed a time-series deep learning model to perform extreme action recognition, built with an efficient dual streams Convolutional Neural Network integrating with Convolutional Long-Short Term Memory. Notably, a novel attempt to employ background-subtracted pose keypoints as input for the recognition. Furthermore, the proposed method demonstrated improved background noise resistance when tested in the datasets of Hockey, Movies, Violent-Flow, and RWF-2000. As a result, the ablation study shows that complementing the RGB frame difference with pose keypoints will improve the framework's accuracy. The performance of the proposed framework is comparable to the existing state-of-the-arts on the RWF-2000 dataset at 87.00% accuracy, 100% accuracy on the Movie dataset, 97.00% accuracy on the Hockey dataset, and Violent-Flows dataset at 92% accuracy. The findings discovered in this study hold enormous potential to advance the current framework of extreme action recognition.

Actions (login required)