Kernel-Based Estimation of Driving Space for Vehicle Hypothesis Generation in Advanced Driver Assistance Systems

Shadow underneath the vehicle is used as a sign pattern to detect moving vehicles in Advanced Driver Assistance Systems (ADAS). Two dynamic thresholds are required to extract the shadow, however, the upper bound is difficult to compute due to the intensity of the shadow depends on illumination conditions. Known methods use the grayscale distribution of the driving space to gain the upper bound with the assumption that the distribution of the road pixels is a normal distribution. However, this assumption does not hold in some real traffic situations such as tree shadows, non-uniform surface, etc., and hence leads to wrong results. For improvement kernel density estimation (KDE) is introduced here to estimate the grayscale distribution of dynamic driving space directly from the data without any assumptions about the underlying distribution. Experimental results show that kernel density estimation can estimate grayscale distributions of any form and detect shadow underneath the vehicle from various real traffic sequences with which known methods fail.