Improved mKLT and low layered HG-CNN based dynamic gesture recognition hardware system

Object tracking in videos is a critical task in computer vision. It comes with challenges due to the processing complexities and the high accuracy requirement. Challenges like varying lighting conditions, partial or complete occlusion, shape changes, and the presence of multiple persons make object tracking particularly difficult. A new dataset named LNMIIT Dynamic Hand Gesture Dataset-5 (numerals 0 to 9) has been prepared under various challenging conditions. An innovative Region of Interest (ROI) hand detection model has been proposed, which utilizes motion and color information to identify hands automatically. The template Matching technique combined with the Improved mKLT (Modified Kanade Lucas Tomasi) tracking algorithm has been used to track the hand. This hybrid approach aims to enhance tracking performance under challenging conditions. Additionally, A novel and robust CNN model named as HG-CNN (Hand Gesture Convolution Neural Network) has been proposed for hand gesture recognition.HG-CNN excels in accuracy and boasts time efficiency, ensuring rapid response times. Additionally, it is engineered to be energy-efficient, making it a compact and resource-sparing solution for real-time applications. The proposed CNN model achieves an impressive recognition accuracy of 99.83%, showcasing its effectiveness in handling object recognition tasks. A comparative study has been carried out with established pre-trained models, namely LeNet5, Inception V3, and VGG16, and has shown the proposed system outperforming in terms of accuracy, time efficiency, and response time.