Local depth patterns for fine-grained activity recognition in depth videos

摘要

Fine-grained activities are human activities involving small objects and small movements. Automatic recognition of such activities can prove useful for many applications, including detailed diarization of meetings and training sessions, assistive human-computer interaction and robotics interfaces. Existing approaches to fine-grained activity recognition typically leverage the combined use of multiple sensors including cameras, RFID tags, gyroscopes and accelerometers borne by the monitored people and target objects. Although effective, the downside of these solutions is that they require minute instrumentation of the environment that is intrusive and hard to scale. To this end, this paper investigates fine-grained activity recognition in a kitchen setting by solely using a depth camera. The primary contribution of this work is an aggregated depth descriptor that effectively captures the shape of the objects and the actors. Experimental results over the challenging “50 Salads” dataset of kitchen activities show an accuracy comparable to that of a state-of-the-art approach based on multiple sensors, thereby validating a less intrusive and more practical way of monitoring fine-grained activities.