Fall detection in multi-camera surveillance videos: experimentations and observations

This paper presents our study on fall detection for ageing care monitoring. We collected a choreographed multi-camera dataset that contains fall actions and other actions such as walking, standing up, sitting down and so forth. In our work, MoSIFT feature is extracted from the videos recorded by each camera. We conduct a series of experiments to show the performance variations of fall detection when different methods are used. We first compare the performance of the standard Bag-of-Words and spatial Bag-of-Words with different codebook sizes. Then, we test different fusion methods which combines the information from the videos recorded by two orthogonally deployed cameras, where a non-linear χ2 kernel Support Vector Machine (SVM) is trained to detect fall actions. In addition, we also use explicit feature maps along with linear kernel for fall detection and compare it to the standard bag of word representation with a non-linear χ2 kernel. Our experiment results show that late fusion of Bag-of-Words with a 1000 centers codebook obtains the best performance. The best result reaches 90.46% in average precision, which in turn may provide a more independent and safer living environment for the elderly.