Vision Based Fall Detector Exploiting Deep Learning

Abstract

In this paper we propose a vision based fall detection algorithm. Our scheme exploits a deep learning paradigm in order to isolate human's object from the background and then to perform tracking. Deep learning better emulates our brain by propagating the raw sensory inputs into "deep" levels of hierarchies. Network adaptation dynamically re-configures the network to fit current environment visual properties. This way, object classification accuracy and tracking is enriched. In the following step, geometrically properties from the detected human object takes place. This is performed by extracting real 3D measurements from the captured 2D image planes. Camera self-calibration methods through the extraction of vanishing points are considered in this context. The derived features are filtered using autoregressive models and filtered feature sequence are exploited as feature in a time delay neural network for performing the final fall detection. Semi-supervised learning strategies are exploited to enhance classification efficiency. Experimental results indicate the efficiency of our proposed algorithm.