Intelligent Vehicle Experimental Test-beds


Human-Centered Driver Assistance Systems


Vision Systems for Occupant Posture Analysis

Vision Systems for Driver Analysis and Interface

Related Research

Stereo and Reflectance Based Occupant Posture Estimation

The feasibility of a stereo based occupant posture estimation system is investigated.  Currently head location and size are relied upon for occupant classification.


The above image shows an example reflectance and stereo disparity pair. The depth information is generated using  left and right rectified and calibrated images that are fed into Small Vision Systems' Stereo API. A background model is first obtained using N frames of disparity data of an empty cabin. After the background image is computed, the current foreground data is generated. A threshold is applied to eliminate noisy pixel stemming from invalid stereo regions. Depths that fall outside the car are also removed. Simple background subtraction is applied and a median filter and morphological opening are performed to eliminate noise. Connected component analysis is also performed to remove areas smaller than a pre-defined minimum head size. This image is the current foreground disparity image.


Given the current foreground image and the raw reflectance image, the best head location can now be located. First edges are found in the reflectance image using Sobel operators. Edges that are contained in the current foreground disparity image are retained. Given the pre-computed set of ellipse templates of various sizes, angles and eccentricities the best fit ellipse/head position is found by maximizing a cost function that takes into account both the ellipse fit and the depth values inside the ellipse.


Position classification for airbag decisions can be made if depth can be obtained from the ellipse location. This requires calibrating the stereo camera based on its position in the real world and converting the detected head coordinate into a real world point. Once obtained, this point, and any other feature points used in future research can be used to make decisions about the occupant posture and location and whether these points correlate to a safe or unsafe airbag deployment decision.


For references, see publications page.