Resumen
In the past few years, it has become increasingly important to automatically detect falls and provide feedback in emergency situations. To meet these demands, fall detection studies have been undertaken using various methods ranging from wearable devices to vision-based methods. However, each method has its own limitations and one common limitation that is prevalent in almost all fall detection studies is that they are restricted to indoor environments. Therefore, we focused on a more dynamic and complex outdoor environment. We used two-dimensional features and Rao-Blackwellized Particle Filtering for human detection and tracking, and extracted three-dimensional features from depth images estimated by the supervised learning method from single input images. As we used the methods in combination, we could distinguish a series of states in which a person falls more precisely and then successfully perform fall detection under dynamic and complex scenes. In this study, we solved the initialization problem, the main constraint of existing tracking studies, by applying the particle swarm optimization method to the human detection system. In addition, we avoided using the background reference image feature for image segmentation due to its vulnerability towards dynamic outdoor changes. The experimental results show a reliable and robust performance for the proposed method and suggest the possibility of effective application to the pre-existing surveillance systems.