A multi-feature tracking algorithm enabling adaptation to context variations

Automatic Parameter Adaptation for Multi-object Tracking

“…The experiments show a significant improvement of the tracking performances while using the proposed controller. Although we test with an appearance-based tracker, other tracker categories can still be controlled by adapting the context definition to Approaches MT (%) PT (%) ML (%) Xing et al [9] 84.3 12.1 3.6 Li et al [10] 84.6 14.0 1.4 Kuo et al [1] 84.6 14.7 0.7 Appearance Tracker [5] without the proposed controller 78.3 16.0 5.7 Appearance Tracker [5] with the proposed controller 85. 5 9.2 5.3 Table 2.…”

“…Approaches MOTA MOTP M Berclaz et al [11] 0.80 0.58 0.69 Shitrit et al [12] 0.81 0.58 0.70 Henriques et al [13] 0.85 0.69 0.77 Appearance Tracker [5] without the proposed controller 0.62 0.63 0.63 Appearance Tracker [5] with the proposed controller 0.87 0.72 0.80 Table 3. Tracking results for the PETS sequence S2.L1, camera view 1, time 12.34.…”

“…Table 2 presents the tracking results of the proposed approach and of some recent trackers from the state of the art. The proposed controller increases significantly the performance of the tracker [5]. The M T value increases 78.3% to 85.5% and the M L value decreases 5.7% to 5.3%.…”

“…In this paper, we select an object appearance-based tracker [5] to test the proposed approach. This tracker takes as input a video stream and a list of objects detected in a predefined temporal window.…”

“…These two approaches require the execution of different trackers in parallel which is expensive in terms of processing time. In [5], the authors propose a tracking algorithm whose parameters can be learned offline for each tracking context. However the authors suppose that the context within a video sequence is fixed over time.…”