A Lightweight Facial Expression Detection Model for Elderly Depression Monitoring on Consumer Edge Devices

Depression is a common and profoundly harmful mental disorder that disproportionately impacts the elderly population. Continuous monitoring of emotional fluctuations in older adults is crucial for both the prevention and treatment of depression. Considering the intrinsic relationship between emotions and facial expressions, this study proposes a lightweight system for depression monitoring specifically designed for elderly individuals. The system included modules for data acquisition and preprocessing, a lightweight architecture, a real-time detection model, and decision-making processes. To address the difficulties associated with deploying models on edge devices and the limited accuracy of facial expression recognition in complex settings, a novel lightweight and efficient detection model is introduced. This model improves detection accuracy while ensuring compatibility between advanced computer vision techniques and hardware with limited computational resources. Experimental findings show that the proposed model reduces parameters, computational burden, and model size by approximately 30%. In addition, it surpasses other contemporary models in detection accuracy, with gains in precision (P) ranging from 1% to 15.2% and a mean average precision (mAP) of 91.5%. Experiments on multiple datasets have also verified its effectiveness and generalization ability. Hardware evaluations on the Jetson Orin Nano development board reveal that the model attains a frame rate of 26 frames per second (FPS) while consuming only 5.551Wh of energy.