Using Depth to Extend Randomised Hough Forests for Object Detection and Localisation

Implicit Shape Models (ISM) have been developed for object detection and localisation in 2-D (RGB) imagery and, to a lesser extent, full 3-D point clouds. Research is ongoing to extend the approach to 2-D imagery having co-registered depth (RGB- D) e.g. from stereoscopy, laser scanning, time-of-flight cameras etc.A popular implementation of the ISM is as a Randomised Forest of classifier trees representing codebooks for use in a Hough Transform voting framework. We present three extensions to the Class-Specific Hough Forest (CSHF) that utilises RGB and co- registered depth imagery acquired via stereoscopic mobile imaging. We demonstrate how depth and RGB information can be combined during training and at detection time. Rather than encoding depth as a new dimension of Hough space (which can increase vote sparsity), depth is used to modify the resulting placement and strength of votes in the original 2-D Hough space. We compare the effect of these depth-based extensions to the unmodified CSHF detection framework evaluated against a challenging new real- world dataset of urban street scenes.