Performance Evaluation of Object Detection and Tracking Systems

This paper presents a set of metrics and algorithms for performance evaluation of object tracking systems. Our emphasis is on wide-ranging, robust metrics which can be used for evaluation purposes without inducing any bias towards the evaluation results. The goal is to report a set of unbiased metrics and to leave the final evaluation of the evaluation process to the research community analyzing the results, keeping the human in the loop. We propose metrics from statistical detection and estimation theory tailored to object detection and tracking tasks using frame-based as well as object-based evaluation paradigms. Object correspondences between multiple ground truth objects to multiple tracker result objects are established from a correspondence matrix. The correspondence matrix is built using three different methods of distance computation between trajectories. Results on PETS 2001 data set are presented in terms of 1 and 2 order statistical descriptors of these metrics.