Combining Multiple Motion Estimates for Vehicle Tracking

of average position error, and in the number of frames where the results outper-form the best instantaneous individual method. These result are encouraging. Acknowledgments We would like to thank researchers from the connectionnist and vision mailing list, Prof. for their help, encouragement and discussions, and for providing image sequences. tive importance of these errors, the vehicle size (in pixels) is reported with each set of measures. In terms of average errors, by looking at table 1, it appears that the estimate of the 1st combining method, although it remains within reasonable bounds, only outper-forms the best individual estimate for one vehicle. The second combining method, however, has much better performance, consistently better than the lowest individual error. In the only case where this is not true, the individual error is already small (1.1 pixels), for a target size of approximately 27 pixels. It appears that the global estimate provided by the combining method 1 performs well only when the features measurement does not present major difficulties for neither individual tracking systems, as predicted in section 3. Apart from this specific weakness, global estimates of both methods are in general sensitive to the following sources of errors: (i) wrong mask ini-tialization; (ii) inaccurate mask update; and (iii) consistently wrong feature measurements in both individual tracking methods (only for the combining method 1). In this paper, we introduced the problem of combining multiple models of a dynamic process, and presented an application to target tracking using multiple motion estima-tors. Two independent tracking systems are used: one is based on the bounding-box of the moving object, the other one uses the object's 2-D pattern. Both individual tracking systems provide motion parameters estimates which are then combined into a global one. Several classes of combining methods presented in the literature are reviewed. Tracking performances are evaluated on 6 vehicles, from three different outdoors image sequences. To precisely compute each individual and global estimate error, the " real " vehicle positions at each frames have been hand entered. The first method is based on a linear combination of the individual estimates, whose weights are inversely proportional to the covariance matrix coefficients. When the two individual methods both provide good estimates, then the results of this combining method represent an improvement, yielding smaller position error than each individual one. When neither individual method performs correctly, however, then combining method does not introduce any improvements. This is …