Camera based fall detection using real-life data

More than thirty percent of persons over 65 years fall at least once a year and are often not able to get up again. Several camera based fall detection algorithms have been proposed in the literature, with the aim of reliably alerting caregivers about older persons' falls at home. These algorithms have been evaluated almost exclusively using brief segments of video data captured in artificial environments under optimal conditions and with falls simulated by actors. By contrast, we collected real-life video data recorded over several months at seven older persons' residences. In this dissertation, we first present an analysis of the real-life video data, including the different falls and the numerous identified challenges. Using this knowledge, we implemented our fall detection system in different phases, tackling known and new challenges. We first built a means to follow the person in the image using a basic background subtraction method. Then we identified the features which are most popular in the state-of-the-art and build a basic fall classifier using a support vector machine. The results from our validation using the real-life data set showed a higher false alarm rate than reported in the literature. We added temporal information to our feature vector for classification to decrease the amount of generated false alarms. The performance of our detection algorithm was compared with that of three previously reported algorithms that used a publicly available simulation data set. All four algorithms produced similar results when using the simulated data. However, the performance of our algorithm degraded drastically when evaluating falls in the real-life data. The false alarm rate was especially high, showing that some challenges still need to be met to make the system sufficiently robust to deploy in real-world situations. We concluded that using more realistic data sets that include longer video recordings and a broad range of activities are essential to reveal weaknesses in fall-detection algorithms. Further we presented several options to tackle the challenges that caused our high false alarm rate. First, the usage of a particle filter combined with a person detector increases the robustness of our foreground segmentation, reducing the number of generated false alarms by 50% and maintaining or even increasing the sensitivity. Second, selecting only non-occluded falls to train the model of our fall detector further decreases the false alarm rate on average from 31.4 to 26 falls per day. But most importantly, this improvement is also shown by the doubling of the area under the curve of the precision-recall curve of this detector compared to a detector using all falls. Third, personalizing the detector by adding several days containing only normal activities of the monitored person to the training data further increases the robustness of our fall detection system. In one case, this reduced the number of false alarms by a factor of 7 while in another one the sensitivity increased by 17% for an increase of the false alarms of 11%. The two final proposals are a bit different: using late fusion to combine multiple cameras and a different approach to detect falls using abnormal inactivity.

Publication year:2017

Accessibility:Closed