High Level Activity Recognition using Low Resolution Wearable VisionSudeep Sundaram, Walterio Mayol-Cuevas, High Level Activity Recognition using Low Resolution Wearable Vision. First Workshop on Egocentric Vision, in conjunction with the International Conference on Computer Vision and Pattern Recognition. June 2009. PDF, 591 Kbytes. External information
This paper presents a system aimed to serve as the enabling platform for a wearable assistant. The method observes manipulations from a wearable camera and classifies activities from roughly stabilized low resolution images (160x120 pixels) with the help of a 3-level Dynamic Bayesian Network and adapted temporal templates. Our motivation is to explore robust but computationally inexpensive visual methods to perform as much activity inference as possible without resorting to more complex object or hand detectors. The description of the method and results obtained are presented, as well as the motivation for further work in the area of wearable visual sensing.