Pre-Impact Fall Detection Based on Multi-Source CNN Ensemble

As the number of aged population grows, fall detection has attracted considerable research attentions in recent years. Through the data collected by wearable sensors and specific algorithms, fall detection and protection can be performed before the user hits the ground. However, most of existing researches have not considered the direction of falls which can be used for more effective protection. In this paper, to further distinguish the direction of falls and improve the detection accuracy, we propose a multi-sensor-based fall detection system by taking the detection as a multi-class problem. To extract the feature from multi-sensor data more effectively, we also present a Multi-source CNN Ensemble (MCNNE) structure. In the proposed system, data from different sensors are preprocessed and formatted as the training dataset independently, and output features map from different sensors are concatenated to construct a overall feature map. Compared with single CNN structure and various ensemble bi-model structures, MCNNE has better performance. On 800 falls and 1000 activities of daily living, the overall average accuracy of our detection system can reach to 99.30%, and false positive rate (FPR) is lower than 0.69%.