Human Action Recognition using Deep Learning Technique

摘要

Identifying human activities from videos is a challenging area of computer vision that requires the recognition of human actions. In this work, we offer a Mediapipe and LSTM based deep learning-based technique for human action recognition. Using Mediapipe, we decode human stance and movement information from videos, and then we feed the information into an LSTM network to categorize the actions. We test our method using a custom dataset that we collected under a number of difficult conditions, such as shifting lighting, shifting camera angles, and occlusion. On our custom dataset, our suggested method performs at the cutting edge with a maximum accuracy of 96.3%. We run comprehensive tests to evaluate how well our method stands up to various scenarios, such as changing lighting conditions, shifting the camera’s angle and occlusion. Potential applications for our suggested methodology include surveillance, sports analysis and human-robot interaction among other areas.