In this paper, we present a video-based method of detecting fall incidents of the elderly living alone. We propose using the measures of humans' heights and occupied areas to distinguish three typical states of humans: standing, sitting, and lying. Two relatively orthogonal views are utilized, in turn, simplifying the estimation of occupied areas as the product of widths of the same person, observed in two cameras. However, the feature estimation based on sizes of silhouettes varies across the viewing window due to the camera perspective. To deal with it, we suggest using Local Empirical Templates (LET) that are defined as the sizes of standing people in local image patches. Two important characteristics of LET are: (1) LET in unknown scenes can be easily extracted by an automatic manner, and (2) by its nature, LET hold the perspective information that can be used for feature normalization. The normalization process is not only to cancel the perspective but also to take the features of standing people as the baselines. We realize that heights of standing people are greater than that of sitting and lying people. People in standing states also occupy smaller areas than whom in sitting and lying states. Thus, three humans' states fall into three separable regions of the proposed feature space, composing of normalized heights and normalized occupied areas. Fall incidents can be inferred from time-series analysis of human state transition. We test the performance of our method on 24 video samples in Multi-view Fall Dataset (1) leading to high detection rates and low false alarms, which outperform the state-of-the-art methods (2) (3) tested on the same benchmark dataset.

No abstract

SUMMARYIn this paper, a fast and automated method of counting pedestrians in crowded areas is proposed along with three contributions. We firstly propose Local Empirical Templates (LET), which are able to outline the foregrounds, typically made by single pedestrians in a scene. LET are extracted by clustering foregrounds of single pedestrians with similar features in silhouettes. This process is done automatically for unknown scenes. Secondly, comparing the size of group foreground made by a group of pedestrians to that of appropriate LET captured in the same image patch with the group foreground produces the density ratio. Because of the local scale normalization between sizes, the density ratio appears to have a bound closely related to the number of pedestrians who induce the group foreground. Finally, to extract the bounds of density ratios for groups of different number of pedestrians, we propose a 3D human models based simulation in which camera viewpoints and pedestrians' proximity are easily manipulated. We collect hundreds of typical occluded-people patterns with distinct degrees of human proximity and under a variety of camera viewpoints. Distributions of density ratios with respect to the number of pedestrians are built based on the computed density ratios of these patterns for extracting density ratio bounds. The simulation is performed in the offline learning phase to extract the bounds from the distributions, which are used to count pedestrians in online settings. We reveal that the bounds seem to be invariant to camera viewpoints and humans' proximity. The performance of our proposed method is evaluated with our collected videos and PETS 2009's datasets. For our collected videos with the resolution of 320x240, our method runs in real-time with good accuracy and frame rate of around 30 fps, and consumes a small amount of computing resources. For PETS 2009's datasets, our proposed method achieves competitive results with other methods tested on the same datasets [1], [2]. key words: local empirical templates, local density ratios, density ratio bounds, and people counting

Abstract-People always make a little contact with the ground during usual activities mainly by feet but often lie completely on the ground after accidental falls. Thus, we propose using Human-Ground Contact Areas (HGCA) to classify human states into standing, sitting and lying states. A fall is defined by a fast change of human states from standing or sitting to lying and continuity in lying state for a sufficient duration. Temporal analyzing human-state transitions is used to discriminate falls from usual events. To measure HGCA, we project foreground of monitored person from one view to another by using homography of the ground between them. Overlap regions between projected foreground and foreground in the latter view that only exist in which people are in contact with the ground, due to plane parallax, are measured as HGCA. We generalize a threshold of HGCA to separate lying states from the others from view-invariant distributions of HGCA with respect to human states. We propose using human state simulation in which camera viewpoints are freely changed to capture 3D human models in various states. Hundreds of images are generated from the simulation as training data to build these distributions. We test our method on "multiple camera fall dataset" leading to a competitive performance with other methods tested on the same dataset.

No abstract