Fusion of Motion and Appearance for Robust People Detection in Cluttered Scenes - Intelligent Video Event Analysis and Understanding - page 121

(a) (b) (c)

(d) (e) (f)

Fig. 8 Examples of human detection in each step. (a) the original slice; (b) initial motion
confidence map only using Gaussian mixture; (c) refined motion confidence map; (d) initial
hypothesises by using pure static human detector; (e) detection results by using the motion
map of (b); (f) refined detection results by using the motion map of (c).

which manually labelled ground truth was available. By varying the threshold of
the detection scores one at a time, we obtain the Receiver Operating Characteristics
(ROC) curves of those detectors, showing false positive rate versus true positive rate
(see Fig 11). When comparing with the ground truth annotation, we measured the
overlap score between the detected bounding box and ground true bounding box.
A detection with overlap score larger than 50% is labeled as a 'match'. We add the
ROC curve of the approach using motion score directly in the Bayesian verifica-
tion process as done in [25]. For further explaining the role of motion information
played in improving the detection rate, we also plotted the ROC curve of a pure mo-
tion detector, i.e. the bounding boxes are weighted only by their motion information
as computed by Eq.(3). From this experiment, it is clear that the motion informa-
tion plays a critical role in accurate detection. Simply using the motion information
alone also gave good detection as shown by the ROC curve. This is because most of
the motions in this particular scene were caused actually by human movement. The
ROC curve shows that our dynamic detector improves significantly the performance
of the static detector by Dalal and Triggs [3] and using the proposed pyramid Gaus-
sian Mixture modeling performs better than just using the motion score. The false
alarms rate has been greatly reduced compared to the static detector. For example,
to achieve a detection rate of 70% on the ticket-office scene, our detector produces
110 false alarms whilst the detector by Dalal and Triggs generated 370 false alarms,
over 3 times more.