Improved activity recognition via Kalman smoothing and multiclass linear discriminant analysis

Improving activity recognition, with special focus on fall-detection, is the subject of this study. We show that Kalman smoothed in-painting of missing pose information and task-specific dimensionality reduction of activity feature vectors leads to significantly improved activity classification performance. We illustrate our findings by applying common classification algorithms to dimensionally reduced feature vectors, and compare our accuracy to previous work. In part two we investigate our methods on a small subset of the data, in order to ascertain what accuracy performance is achievable with the smallest amount of information available.