Object Tracking over Multiple Uncalibrated Cameras Using Visual, Spatial and Temporal Similarities - Advanced Concepts for Intelligent Vision Systems - page 175

The values used for each parameter of the STAC algorithm during evaluation
were: d s max =2 . 0; w t min =1 . 0; m min = 20. These values were selected empiri-
cally, and based on experience with our particular signature type. Additionally,
a maximum of 100 foreign track snapshots were stored by a camera at any time.

3.3 Evaluation Metrics

We considered two metrics for our evaluation. The F 1 score measures the ro-
bustness of results. The correspondence delay (CD) measures how long it takes
to link tracks. A higher F 1 score and a lower CD are desirable.
The F 1 score is the harmonic mean of recall and precision, balancing the
trade-off between false positives and false negatives. Recall and precision have
been used previously in evaluating tracking-based events [8]. Here, we treat a
correct correspondence as a true positive, an incorrect correspondence as a false
positive and the lack of a correspondence where one should have been found as
a false negative. The F 1 score is given by,

F 1 =2 precision

recall
precision + recall .

·

(13)

For a given correspondence, the correspondence delay (CD) is the time between
the second track becoming visible and a correspondence being found between
the two tracks. The CD for a test set is the mean of the CDs for the true
positive correspondences for that test set. To our knowledge, the timeliness of
correspondences has not been considered in the literature before.

4 Results and Discussion

Table 1 shows a summary of the results of evaluating the STAC algorithm across
all test sets for the two cases of:

- the STAC algorithm described in section 2 but without using the location-
based kernel matching method described in section 2.1; and
- the STAC algorithm with the use of location-based kernel matching.

Table 1. Overall results of evaluating the STAC algorithm with and without location-
based kernel matching (LKM) across all seven test sets

Metric Without LKM With LKM

Recall 0.54 0.66

Precision 0.63 0.68

F 1 score 0.58 0.67

CD (frames) 31.2 21.7