next up previous
Next: Bibliography Up: icra02HTML Previous: Experiment 2.

Summary and Conclusions

We have developed a visual tracking system where the consensus of simple visual cues facilitates both robust and real-time performance. Response and action fusion in a voting based framework are proposed. The approaches differ in the choice of the underlying voting spaces. Special emphasis was put on the evaluation of the cues' weights where four methods have been used. The experimental evaluation considered the effect of the weighting technique to the performance of the fusion approaches. It has been shown that the most accurate and reliable results are obtained using the texture based weighting where the weights are set in advance and kept constant during tracking. The system was also evaluated with three most common camera-object configurations: i) static sensor/moving object, ii) moving sensor/static object and iii) moving sensor/moving object. Experimental evaluation has shown that the best performance was obtained using the response fusion approach. On 30 test runs this approach was successful in 90% of cases compared to 73.7% for the action fusion approach. To demonstrate the necessity for the fusion, color, differencing and correlation were also evaluated individually. It has been shown that color performed best. Color is commonly used in cases where the object is uniform in color and, given some assumptions about the environment, it should be considered as a strong tracking candidate. The results are presented for a monocular system and our future work will consider binocular cues in the fusion process.

next up previous
Next: Bibliography Up: icra02HTML Previous: Experiment 2.
Danica Kragic 2002-12-06