Real-Time Suspicious Human Action Recognition from Surveillance Videos for Resource-Constrained Devices
I. Subha, P. Narmadha, S Nivedha, T. Sethukarasi · Journal of Computational and Theoretical Nanoscience · 2020
Recent developments in computer vision are seen as a vital advancement in video surveillance. The goal of this research is to build a deep learning model that is capable of analyzing and classifying the video from running CCTV streams to detect criminal actions and identify suspects on the scene. In particular, we focus on the detection of dangerous human behaviors in surveillance videos. This work provides a low cost embedded solution that can be integrated with the existing CCTV cameras. This integration can reduce the cost of transmitting the data to any centralized server, which may have various privacy implications and takes much inference time. We also benchmark our models performance with the existing real-world dataset in terms of accuracy and resource constraints. Using the concept of Multiple Instance Learning on the histogram of the optical flow of the videos combined with the pose estimation of the persons on scene, we provide a lightweight model which has 13 times lesser inference time than the existing very deep models. Focusing on one important thing, this research will expand to which state-of-the-art deep neural networks will “see” violence in photographs and videos, and recognize criminal behavior using characteristics such as gestures, gait, and unethical behavior. This helps enforcement agencies to unravel crime cases faster and also to scale back crimes by identifying the suspects in the surveillance videos.