REAL-TIME HUMAN MOTION DETECTION AND ACTIVITY RECOGNITION

Abstract

One of the interest areas of computer vision is real-time human motion detection, tracking, and activity recognition. It has many applications in a variety of fields, including video processing, abnormally detection, behavior prediction, human- computer interaction, video surveillance, and content-based image retrieval systems. This technology is essential in the fight against crime, terrorism, and threats to public safety. Due to variations in human appearance, changes in illumination, and the volume of data generated, video-based real-time human activity recognition is a difficult and demanding task. Supporting a safe and secure environment for real-time motion detection, tracking, and activity recognition is the aim of this research. The system detects human body parts with skeleton and to define activity based on joint sequence movement and to extract more reliable manner for overlapping area and to solve similar pose with different activities. The goal of this proposed system is to enhance an automated video surveillance system that can identify and track people in both indoor and outdoor settings. The main step of the system involves motion detection, tracking and activity recognition through several steps: First, the system is designed to capture input video and extract region of interest for each frame. And generate features to estimate human and to detect 2D joint projected positions. Then, human detection is applied by using OpenPose detector and categorizes 2D joint sequence of body parts. The system recreates a human skeleton joint in three dimensions using spatial-temporal integration of human body parts. Finally, recognizes the activities such as standing, walking, sitting and running according to joint collection distance and displacement of skeleton joint position. With a deep learning framework, the proposed method operates a robust human skeleton model that is unaffected by changes in the environment or various circumstances. Using joint estimation and position recognition, the system builds a skeleton model from the data perception. The objective of this research is more robust and efficient approach in human detection and activity recognition system from training and testing of multiple data generation by using deep learning approach to recognize different human activities changes in real life environment. The system's total accuracy is 94%, and the proposed approach performs better than expected when it comes to 3D skeleton model-based human detection and activity recognition.