Gesture recognition based on depth difference distribution

We present a novel motion descriptor for gesture recognition based on depth camera. Since each object motion leads to a specific depth change characterized by depth difference, we can recognize object motion via Depth Difference Distribution (DDD) in object region. The DDD is approximated by DDD descriptor in three steps. First, each pixel´s depth difference value is quantified into Depth Difference (DD) codes. Second, the object region is separate into several sub-regions. Third, in each sub-region, a vector is generated to describe the distribution of each DD code. The vectors of each DD code in each sub-region are cascaded into the final DDD descriptor to approximate the depth difference distribution caused by object motion. DDD descriptor is a combination of both motion and shape information. Experiment shows a robust gesture detection performance is achieved within large distance and view angle variation range.