Performance enhancement of vision based fall detection using ensemble of machine learning model

An automatic fall detection system (FDS) provides timely medical assistance to the elderly by predicting falls and non-falls, thereby preventing serious injuries and death. This article proposes an ensemble-based FDS focuses on improving the fall detection accuracy by selecting optimal classifiers with a greedy algorithm-based majority voting approach. The proposed ensemble learning is better than any of the individual models because it involves multiple machine learning models along with deep learning model, namely Support Vector Machine, K-Nearest Neighbour, Decision Tree, and Deep LSTM that are combined in majority voting fashion. The greedy algorithm-based majority voting uses three searching criteria, namely forward search, backward search, and recovery search to select the optimal classifiers. The forward search fuses the classifiers based on majority voting error (MVE) and goodness of classifiers, while the backward search removes the classifier based on the majority voting improvement selection criteria. Thus, the classifiers with the lowest MVE and processing time are fused, and the others are removed. Also, the recovery search is introduced to prevent any loss of optimal classifiers by backward search. Finally, classifiers with the best performance are combined to produce accurate fall detection. Then simulation is carried out for the individual ML model, and the proposed ensemble model. According to the evaluation outcome, the proposed ensemble-based FDS achieves high accuracy (99.98%), sensitivity (99.8%), specificity (99.9%), and precision (96.23%) than the conventional approaches. Hence, the proposed ensemble-based FDS proved to enhance fall detection accuracy.