On-line fall detection via a boosted cascade of hybrid features

In this paper, we present a cascaded learning approach for online fall detection in streaming videos. In the cascaded approach, we propose using hybrid features ranging from Haar-like features to motion boundary histogram (MBH) rather than involving only one single type of features as in traditional cascaded methods. The simple features are employed at earlier stages to rapidly filter out the majority of none fall video clips which are relatively easy to detect, while complex features are used incrementally at later stages to handle samples that are difficult to distinguish. Similarly, the classifier at each stage is trained from simple to complex based on the gradient boosted tree. We further conduct a fall dataset with 400+ hours indoor videos to provide enough samples for the training. Experimental results show that our proposed fall detector achieves significantly higher performance than state-of-the-art approaches in term of both accuracy and speed.