Fall Event Detection System using Inception Densenet Inspired Sparse Siamese Network

A novel few-shot Siamese architecture inspired by Inception and Densenet architectures is proposed as a fall event detection system to detect fall events in signals obtained from waist-worn inertial measurement unit sensors. The proposed system consists of an Inception module followed by a relatively sparse Densenet-based module on each arm of the Siamese network to effectively learn feature representations for the detection of fall events. The proposed system is tested using the SisFall dataset. The proposed system's performance in a few-shot scenario is compared with fall detection systems based on the regular Inception and Densenet121 architectures and the state-of-the-art Siamese convolutional autoencoders. The proposed system outperforms all three fall detection systems. The proposed fall detection system achieved F-scores of 97 ±4% and 68.5 ±10% in 15-shot and 1-shot learning scenarios, respectively.