A Lightweight DualStream-BioInceptNet CNN for Real-Time Suspect Identification Using Face and Gait Recognition

Biometric recognition systems are critical to the security and surveillance use case, but they often have challenges in the real world, especially with occluded or poorly lit faces. In such situations, gait-based recognition is a suitable alternative. This research was inspired by how humans identify others naturally through facially identified features or gait/gestures. Based on this natural composition of human recognition strategies, a dual modality model incorporating a face and gait recognition was developed. The proposed lightweight CNN architecture called DualStream-BioInceptNet, which combines inception modules with compact dense layers to process each frame to extract spatial information of individuals with efficacy. The model was purposely designed to avoid computationally heavy or recurrent layers, which allows implementation at a level that would provide both efficiency and real-time capability, including technology with poor computing capabilities as needed in modern biometrics under constrained conditions. From the evaluation of our model on the customised dataset with variable real-world conditions, a recognition accuracy of 97.92% was achieved, outperforming both single-stream approaches such as face-only stream (96.83%) and gait-only stream (95.65%). It operates efficiently with only 12.4 million parameters and 2.88 GFLOPs, making it suitable for real-time biometric recognition applications.