Demo of “Counting People using Low-level Features and Bayesian Regression”

These are demo videos for the TIP journal paper, “Counting People using Low-Level Features and Bayesian Regression.” The video is available in Quicktime format (H.264). The datasets are available here.


Peds1 Dataset

This are the video results for the Peds1 dataset, which contains 3 motion classes (away, towards, and scene). The red and green segments are the “away” and “towards” components of the crowd. The estimated crowd count for each segment is shown in the top-left, with the (uncertainty) and the [ground-truth]. The prediction for the “scene” class, which is the total number of people in the scene, is shown in the top-right in blue. The ROI is also highlighted.

Results on 4000 frames

These are results on the 4000 frame with ground-truth counts. The system was trained on 1200 frames.

[mov (16 MB)]

These are the plots of the counts over time for the 4000 frames. Gray levels indicate probabilities of the predictive distribution.

away
towards
scene

Results on full video

The system was trained on 2000 frames with ground-truth, and tested on the remaining 50 minutes of video.

Training Video
[mov (16 MB)]
Test Video
[mov (5.0 MB)]
[mov (13 MB)]
[mov (10 MB)]
[mov (9 MB)]
[mov (12 MB)]
[mov (8 MB)]
Count Plot – over the whole 55 minute video.
away
towards
scene

Peds2 Dataset

This is the video results for the Peds2 dataset, which contains 5 motion classes (left-slow, right-slow, left-fast, right-fast, and scene). The red and green segments are the “right-slow” and “left-slow” components of the crowd, the blue and yellow segments are the “right-fast” and “left-fast” ones. The estimated crowd count for each segment is shown in the top-left, with the (uncertainty) and the [ground-truth]. The count for the “scene” class, which is the total number in the scene, is shown in white text.

Results on 4000 frames

These are results on the 4000 frame with ground-truth counts. The system was trained on 1000 frames.

[mov (16 MB)]

These are the plots of the counts over time for the 4000 frames. Gray levels indicate probabilities of the predictive distribution.

right-slow
left-slow
right-fast
left-fast
scene

Results on full video

The system was trained on 2000 frames with ground-truth, and tested on the remaining 50 minutes of video.

Training Video
[mov (15 MB)]
Test Video
[mov (4.5 MB)]
[mov (10 MB)]
[mov (9 MB)]
[mov (6 MB)]
[mov (8 MB)]
[mov (5 MB)]
Count Plot – over the whole 55 minutes of video
right-slow
left-slow
right-fast
left-fast
scene