Enhanced Robustness in Low-Cost Doppler Radar Based Fall Detection System via Kalman Filter Tracking and Transition Activity Analysis

Fall detection systems based on low-cost Doppler radars often encounter challenges due to false alarms, particularly during transition activities such as sit-to-stand and stand-to-sit motions, sudden entrance into the sensor region, and abrupt exits from the sensor region. This paper aims to address this issue by reducing false alarms and enhancing the system’s robustness by integrating different transition activities alongside routine activities, utilizing a 3.18 GHz low-cost Doppler radar. The acquired radar data is categorized into stationary, transitions, and falls. A Kalman filter of the constant jerk model is employed to track all activities, enabling the extraction of features from higher-order motion parameters like acceleration and jerk. Two classification approaches are utilized: binary classification (fall vs. other activities) and multiclass classification (stationary vs. transitions and fall). Various machine learning algorithms are applied, and the results indicate that the Decision Tree algorithm achieves the highest accuracy of 85.71% for multiclass classification, while the k-Nearest Neighbors (kNN) algorithm yields the best performance with an accuracy of 92.85% for binary classification. We conduct a comparative study between our work and recent state-of-the-art approaches and find that our results are competitive.