•  
  •  
 

Abstract

Hand gesture recognition is a challenging problem in computer vision, particularly in terms of security surveillance applications. This study presents the first efficient system for abduction-related hand gesture real-time detection based on deep learning. The most critical problem is to detect and recognize hand gestures in real surveillance conditions and to be computationally effective for real-time multi-hand tracking in various lighting situations while allowing reliable surveillance beyond the 1–4 meters limitation. The proposed system consists of three main parts: The adaptive hand tracking algorithm, which has been used to create the Abductees-Rescue dataset. Introduced pose estimation You Only Look Once version eight, the MediaPipe Hands framework, which variable 3D landmarks extraction at distance. Adaptive Long Short-Term Memory, which was produced by an optimized LSTM model. It has been implemented with the help of the multi-threaded notification approach. All three parts are combined to create novel approaches to the field of abduction rescue research. The Abductees-Rescue dataset includes 9,111 samples, 4,545 native and 4,566 abduction-related 3D hand landmarks. The optimized LSTM model multiplies into a hierarchical signal flow with L2-regularization and an additional dropout approach of achieving 96.12% classifying efficiency. Multi-threaded notification has been implemented, allowing for more than two hand-tracking at the same time, all for being 15 FPS sometimes. Pose estimation has been adapted to temporal detection that allows 10-meter-radius action from perpetrators. It utilizes the MediaPipe Hands framework adapted to 3D landmark actions in the 10 meters distance. Integrated with the Telegram Bot API to enable the multi-threaded engineers to send messages with a picture of abduction delivery time, 1.839 seconds delivery time. This system is a significant advance in building surveillance-based gesture recognition for assurance.

Download

Share

COinS