Project 3: Test Videos

These videos have been selected as exemplary of a range of situations and imaging conditions. You should test your trackers on a variety of different types of sequences. Do not only test them on the car sequences, for example, as these are meant for calibration purposes only.

These videos are relatively simple examples taken under laboratory conditions. They contain a single moving target: a red remote-controlled car. The videos are taken from a number of viewpoints and are specifically designed so that background subtraction can work well on them.

1. Overhead view:
   • car-overhead-1.avi
   • car-overhead-1-hires.avi
   • car-overhead-2.avi
   • car-overhead-2-hires.avi
   • car-overhead-3.avi
   • car-overhead-3-hires.avi
2. Perspective view:
   • car-perspective-1.avi
   • car-perspective-1-hires.avi
   • car-perspective-2.avi
   • car-perspective-2-hires.avi
   • car-perspective-3.avi
   • car-perspective-3-hires.avi
   (note that the last two are challenging because, as the student who took these videos comments, "Some idiot is moving around in the background.")

These videos are standard sequences used for performance evaluation of video surveillance algorithms and applications. They typically have many moving targets and complex background and/or illumination changes. They are also very high resolution video. Feel free to downsample the video as you like for evaluating trackers on these sequences. Useful tools for this are avidemux under Linux and virtualdub for Windows.

See the corresponding pages for detailed information about each sequence, as well as ground truth tracking information for each.

1. HERMES Indoor Sequences

   In this HERMES sequence (2003 frames @ 15 fps, 1392 x 1040 pixels) there are three people in a room. They act in a discussion sequence sitting around a table, where bags are carried, left and picked from the floor, and bottles are carried, left and picked from the vending machine and from the table. In this discussion sequence several agents are involved in different simultaneous grouping, grouped and splitting events, while they are partially or completely occluded.

2. HERMES Outdoor Sequences

   This HERMES sequence (1612 frames @ 15 fps, 1392 x 1040 pixels) presents a great diversity of situations. Three people and three cars act on a robbery sequence, where suitcases and bags are carried, left and picked from the floor. Multiple interaction events can be seen, in which several agents are involved in different simultaneous grouping, grouped and splitting events, while they are partially or completely occluded. Among the sequence difficulties, we must also remark that objects from the initial background are removed, several targets suffered from heavy background camouflage, and the strong clutter caused by similar group partners.

3. Zebra Sequence

   A crosswalk sequence is analysed in this database (1344 frames @ 25fps, 720 x 576 pixels). Four people are involved in different interaction events. Further, several vehicles cross the scene in a front plane, as people walk behind various streetlamps and trees, resulting in multiple partial and complete occlusions of the targets.

4. VISOR Outdoor Sequence

   This VISOR sequence (2292 frames @ 10 fps, 384 x 288 pixels) presents fifteen people walking on square. People interact with each other, they get involved in different simultaneous groupings and have different paths. Also, there are people walking behind several columns, which results in multiple occlusions (partial and complete) of the targets.