Abstract
This paper provides a deep learning-based gesture analysis framework for real-time crime prediction to overcome the shortcomings of current criminal detection systems. Advanced neural networks like Convolutional Neural Networks (CNNs) are used for the detection of anger, stress, and aggression, and LSTM or Long Short-Term Memory along with Graph Convolutional Network (GCN) models are used for audio-emotion detection associated with suspicious or criminal behaviour. By deploying deep learning models like Edge AI on edge devices for on-the-spot analysis and OpenPose or MediaPipe for tracking gesture movements. The proposed solution combines gesture recognition with pose estimation, facial micro-expression, and audio-emotion detection, making it a strong analysis of non-verbal cues in crime investigation. This approach will make the currently available surveillance systems more efficient and provide timely alerts to improve public safety.