Optimized Multimodal Human Activity and Anomaly Detection

Abstract

Human activity recognition and anomaly detection play a crucial role in applications such as intelligent surveillance, healthcare monitoring, and smart environments. This study proposes a optimized multimodal framework that integrates video data and IMU sensor signals to differentiate normal and abnormal human activities efficiently. Each data stream is processed independently using optimized, low-complexity deep learning models, and predictions are combined at the decision level to enhance accuracy while avoiding the overhead of feature fusion. Model Optimization reduces memory usage, model size, and inference time, enabling deployment on edge devices such as smart cameras, smartwatches, and smartphones. The system is evaluated on various activities, including fall, driver drowsiness as well as normal activity like walk, run etc, under diverse lighting conditions, sensor placements, and environmental variations to ensure robust performance. The framework emphasizes real-time monitoring and low-latency response, providing a scalable and practical solution for continuous anomaly detection. Future work may extend the system to additional modalities and incremental learning for improved adaptability.