PETS2010 and PETS2009 Evaluation of Results Using Individual Ground Truthed Single Views

Challenges of Ground Truth Evaluation of Multi-target Tracking

“…In such case, the aim is to achieve the highest possible precision and recall while less attention is paid to the number of ID switches. This may also be the motivation of [11], who impose a logarithmic weight on the number of mismatch errors when computing the MOTA score.…”

“…The only currently existing method (that we are aware of) to objectively measure the performance of a tracking algorithm is to send the results on the S2L1 sequence (represented by bounding boxes) to the PETS organizers [11]. The computed CLEAR MOT metrics, evaluated with respect to unpublished ground truth, are then sent back to the authors.…”

Multiple Target Tracking with Motion Priors

“…Last, in Fig. 4, we have compared the results of our own tracking strategies with the results obtained by several other authors as reported in [12]. For all indicators, the obtained results are quite competitive, and as we can see the most important result is the ATA metric (green column) which reflect the continuity of trajectories, in other words we can follow all targets using less trackers as other approaches.…”

“…Since ground truth data is not available, we have manually generated one considering a rectangle around each target in the sequence. The two flavours of the particle filter we propose here (that differ by using or not the motion priors in the probabilistic motion model) are evaluated with a common methodology that has been used to evaluate the trackers performance [12]. Moreover, we also evaluate how the process of filling incomplete prior data improves tracking results.…”

“…1. We computed the regularized velocity orientation map using > 3000-frames from different sequences of the PETS'2009 dataset [12]. The left image is an example of a histogram for a single pixel, while the right image shows the main two local maxima for the orientation histogram, at each pixel.…”