Performance Prediction Analysis of a Point Feature Tracker Based on Different Motion Models

This paper provides performance prediction analysis techniques for a linear point feature tracking algorithm based on different motion models. We provide closed-form expressions for evaluating the probability of correct data association of a tracker (analyzed with different motion models), when tracking under clutter. We also extend our analysis for the prediction of correct data association when a tracker recovers from a false match to regain correct tracking. The simple mathematical expressions provided here can be used to implement performance analysis procedures that are fast, easy, and reasonably accurate (compared with conventional computationally expensive Monte Carlo tracking experiments employed to predict the performance of a tracker). We have also demonstrated the importance of using a correct motion model for a visual tracker to get optimum tracking performance, based on empirical evaluation techniques. The performance of a trackerʹs robustness under varied noise has also been investigated.