Detecting Human Activity in Thermal Low-Light Videos using Score-Level Fusion of EfficientNet and MobileNet

Introduction: Human Activity Recognition (HAR) in low-light environments is indispensable for healthcare, surveillance, and assisted living, as visibility is limited. The accuracy remains a challenge in Thermal imaging enables activity detection in such conditions. This work introduces a lightweight score-level fusion framework via the Weighted Quadratic Average Method (WQAM) in combining MobileNetV2 and EfficientNetB0 to boost recognition performance on thermal video data. Methods: A dataset of 1,000 low-light thermal recordings covering four events, such as walking, running, duck walking, and crawling, was collected using a FLIR C5 imager. Frame extraction and normalization involved in preprocessing. To improve classification reliability while maintaining efficiency, MobileNetV2 and EfficientNetB0 were trained individually, and WQAM is used to fuse their outputs. Results: MobileNetV2 achieved 87% accuracy, and EfficientNetB0 reached 90%. The proposed fusion model outperform both individual models and state-of-the-art methods significantly and improves performance, achieving 96% accuracy. The approach also demonstrated a reduced processing time suitable. Discussion: The complementary features of MobileNetV2 and EfficientNetB0 are leveraged in WQAM-based fusion effectively, providing computational efficiency and superior recognition accuracy. For real-world applications where accuracy and speed are equally critical, this framework provide suitable balance. Conclusion: The proposed WQAM fusion framework surpasses baseline models and existing methods, achieved 96% accuracy on low-light thermal HAR. This fusion provides a practical solution for deployment in surveillance, healthcare, and assisted living environments with robustness, efficiency, and scalability.