New AI blends CNNs and LSTMs for faster violence detection

A new deep learning architecture achieves 93% accuracy in identifying violent actions within video footage, exceeding the performance of many existing systems. Birendra Kumar Verma of Bennett University and colleagues combined Time-Distributed Convolutional Neural Networks with Long Short-Term Memory networks and a spatial attention mechanism to improve both speed and precision. The resulting model is computationally lightweight, making it suitable for near-real-time deployment in surveillance applications where identifying critical events is paramount. This work demonstrates that hierarchical integration of convolutional and recurrent layers with attention significantly improves recognition accuracy.

TD-CNN and LSTM Architecture for Real-time Violence Detection

Achieving 93% accuracy on the Hockey Fight Dataset, a new deep learning architecture demonstrates significant advancement in automated violence detection within video footage. The core innovation lies in the integration of Time-Distributed Convolutional Neural Networks (TD-CNNs) with Long Short-Term Memory (LSTM) networks, a combination not previously demonstrated with this level of success. This hybrid approach incorporates a spatial attention mechanism positioned between convolutional layers, which adaptively emphasizes crucial spatial areas within each frame, enhancing the model’s ability to differentiate relevant features.

The LSTM component then analyzes the sequential relationships between frames, capturing the temporal dynamics indicative of violent actions. This layered processing allows the architecture to recognize violence at multiple levels, from individual aggressive gestures to complex interactions unfolding over time. Despite its high performance, the architecture is designed to be computationally lightweight, a critical feature for deployment in real-world surveillance systems.

Researchers specifically targeted a balance between accuracy and speed, recognizing that rapid identification is paramount in time-sensitive security applications. The team tested the model extensively, confirming its suitability for near-real-time processing without sacrificing the precision needed to minimize false alarms. Manipal Academy of Higher Education provided open access funding for this research, supporting the development of a system that could improve public safety and security monitoring capabilities. The authors declare no competing interests, and the complete findings are available for review.