Video and Inertial Sensors Fusion for Activity Recognition: A Machine Learning Approach

Human activity tracking analyzes mobility using data from several sensors. This paper presents a multisensor model for fall detection and activity recognition utilizing inertial and RGB data. Inertial data was filtered with a Kalman filter for smoothness, extracting characteristics like GMM and Parseval’s energy, while RGB data was preprocessed using a bilateral filter and gave features like geometric features, full-body curve and full-body ridges. Results were fused with multimodal survey fusion, optimized with Naive Bayes, then classified using AdaBoost. The proposed model was tested on the UR Fall Detection (URFD) dataset, the model achieved 89% accuracy.