Evaluation of Human Detection Algorithms in Image Sequences - Advanced Concepts for Intelligent Vision Systems - page 126

Fig. 3. Extract of a scenario example defined by the industrial partners involved in the
CAPTHOM project

- Set 1: scenarios involving a normal use of a room. In these scenarios, we need

to detect humans that are static or moving, sitting or standing in oces,
meeting rooms, corridors and dining rooms.

- Set 2: scenarios of unusual activities (slow or fast falls, abnormal agitation).

- Set 3: scenarios gathering all false detections stimuli (illumination variation,

moving objects etc).

In the following, Set 4 is defined as the union of these 3 sets. In total, we used
29 images sequences in 10 different places. Images have a resolution of 320 x 240
and have an ”average” quality. Each images sequence lasts from 2 to 10 minutes.
Figures 4 and 5 present results obtained with the CAPTHOM algorithm on
videos extracted from our test dataset.

Fig. 4. Example of results obtained with the CAPTHOM method on a video present-

ing partial occlusion

The choice of the evaluation metric parameters, done for this study, corre-
sponds to an expected interpretation compromise which can be encountered in
many real applications. We use a parameter α , set at 0.8, to balance the local-
ization and the recognition scores. This high value has been chosen to maintain
an important weight for the penalization of bad localization. It results from a
wide subjective evaluation of interpretation results we conducted, involving re-
searchers of the French community, to better understand when a bad localization