Precise Motion Descriptors Extraction from stereoscopic footage using DaVinci DM6446

A novel approach to extract target motion descriptors in multi-camera video surveillance systems is presented. Using two static surveillance cameras with partially overlapped field of view (FOV), control points (unique points from each camera) are identified in regions of interest (ROI) from both cameras footage. The control points within the ROI are matched for correspondence and a meshed Euclidean distance based signature is computed. A depth map is estimated using disparity of each control pair and the ROI is graded into number of regions with the help of relative depth information of the control points. The graded regions of different depths will help calculate accurately the pace of the moving target and also its 3D location. The advantage of estimating a depth map for background static control points over depth map of the target itself is its accuracy and robustness to outliers. The performance of the algorithm is evaluated in the paper using several test sequences. Implementation issues of the algorithm onto the TI DaVinci DM6446 platform are considered in the paper.