Abstract: The static or moving shadows in traffic scenes tend to reduce the accuracy of vehicle target tracking.Thus,it is an important step to effectively remove the shadows in the processing of traffic surveillance videos.However,there hardly is an efficient shadow removal method yet at present,which resists both static and moving shadows by simultaneously exploring the spatial and frequency characteristics of shadows.Under such a circumstance,this study proposed a shadow removal method for traffic video by using a joint voting strategy in spatial-frequency domain.The surveillance video is converted from RGB space into HSV space and then performed non-subsampled shearlet transform (NSST).Assuming that NSST coefficients follows the Gaussian distribution,the mean and standard deviation of coefficients is used to compute the weighted mask for each scale.Subsequently,the weighted mask at coarse scale is employed to adjust the mask at fine scale,according to the zerotree characteristics of multiscale coefficients.The weighted masks of different scales and color channels are thus linearly combined to form a unified mask,which is then binarized by an adaptive threshold calculated by the maximum entropy segmentation based on the least square method.Finally,the moving vehicle area after shadow removal is determined by a joint voting strategy by using the weighted frequency-domain mask,the S-channel mask and the V-channel mask respectively.Experimental results show that the proposed algorithm can effectively remove the static and moving shadows in traffic surveillance video.It reduces the average Euclidean distance by 95% between the ideal trajectory and the output vehicle trajectory of traditional mean shift algorithm,suppressing the interfe-rence of shadows.Meanwhile,the proposed algorithm enhances the robustness of the intelligent analysis and avoids the phenomenon of losing the target.Our research indicates that it is conducive to obtaining more accurate shadow removal result to effectively combine the representation of traffic surveillance video in spatial and frequency domains,and to fully explore the differences of texture features and color features between moving vehicle areas and shadow areas.The accuracy of vehicle target tracking will be therefore improved.

Key words: Joint voting, Multiscale weighted mask, Non-subsampled Shearlet transform, Shadow removal, Traffic surveillance video