Linear depth reconstruction for RGBD sensors

Consumer-level depth cameras, referred to as RGBD devices, are important components to robotic recognition, mapping and navigation systems. Past research has provided detailed models for measurement and quantization noise inherent to these devices. Yet, to date, there has been no published work showing how to leverage these noise models to reduce the measured depth error. Existing approaches rely on colored images registered to the depth image to reconstruct depth, which work best when the device is calibrated and the scene lighting and surfaces allow for a Lambertian model. The proposed method is attractive since it works directly on the depth data without any need for calibration or assumptions regarding the scene. Reconstruction is accomplished using a two stage filter. The first stage removes impulse, i.e., “salt-and-pepper”, noise and adds noise, i.e., dithers, the depth at quantization boundaries. The second stage low-pass filters the depth to remove the added dithering noise. The dithering process is particularly useful for quickly removing large errors in depth at the extreme of the device range where the depth quantization and impulse noise incurs significant error. The proposed reconstruction approach has linear computational complexity and low computational cost. The algorithm is particularly useful for extracting smooth surfaces at the upper limits of the sensor measurement range where impulse noise and quantization errors are large (~7.5cm) and can significantly degrade the performance of downstream recognition, navigation, and mapping algorithms.