Use of an Evaluation and Diagnosis Method to Improve Tracking Performances

This paper presents a general framework for analyzing the evaluation of a tracking algorithm in order to improve it. We first propose a classification of the various errors encountered during the motion detection and the tracking process. This classification is done using a comparison between tracking outputs and ground truth. We propose two evaluation algorithms, a global one and a more precise one. Second, we show how to use this classification to diagnose the tracking errors and to find relevant parameters to solve each problem type and to determine criteria to tune these parameters with respect to the scene environment. This technique is applied to the tracker module of a video interpretation platform whose main goal is to recognize human behaviours. Results are presented for several video sequences taken from a static calibrated camera in three different contexts: a bank agency, a metro platform and an office