Qualitative and Quantitative Evaluation of the Reading People Tracker

People tracking is one major part of visual surveillance systems. A number of different people trackers exist nowadays. One of those multi-people trackers is the Reading People Tracker. It consists of four individual co-operating tracking modules. Weaknesses of one module are thereby compensated by the strengths of another one. The Motion Detector performs background-foreground segmentation to extract moving regions. The Region Tracker tracks the regions from the Motion Detector over time. Region merging and splitting helps to overcome problems during motion detection. The Human Feature Detector detects possible head positions in moving regions. But the fundamental tracking module of the Reading People Tracker is the Active Shape Tracker. It uses a 2D appearance model of human beings, the Active Shape Model. A person’s outlines are represented by an Active Shape which is actually a closed parameterized B-spline curve. An additional feature of the Reading People Tracker is its use of camera calibration to perform image measurements. This diploma thesis evaluates qualitatively and quantitatively the Reading People Tracker. In order to perform the evaluation, a large set of test video sequences is captured. A capturing and calibration environment is implemented to rapidly capture video sequences. Camera calibration is accomplished using Tsai’s perspective pin hole camera model. A particular method is explained to eliminate the effects of radial lens distortion. Input image points to the calibration algorithm are computed in sub-pixel accuracy by fitting lines to a checkerboard calibration target. The focus of the evaluation part is put on the Motion Detector and the Active Shape Tracker. Evaluations are accomplished by comparing tracking output objects with a set of ground truth objects according to different performance metrics. These metrics are pixel-based as well as object-based. The ViPER package is used to annotate test video sequences and to compute the evaluation performance metrics.