If video does not load, please save it to local disk.
- Calibrating PTZ cameras using a image-based
technique described in paper.
- Illumination invariant background subtraction using
two cameras. We show in this project how two cameras can be
used to improve tolerance to specularity, noise and detection
of near-background objects (a common problem with
disparity-based detection methods). Video-rate performance is
also maintained without the use of any specialized
hardware.(paper)
- Headlight sequence
- Indoor sequence
- Near background sequence
- Daytime raining (specularity) sequence
- Nighttime (specularity) sequence
- The use
of two cameras to solve the problem of background subtraction
can be extended to multiple cameras. Specifically, we use
selective stereo to compliment counting the number of people
in crowded scene by segmenting their shadows on the ground
plane. We also ask the question of how to improve results via
a probabilistic selection of cameras that are to be utilized
in the collection of multi-view
images.
- Multiple camera detection and
tracking under occlusions.(may
need to save to your machine before playing
it)
- An active camera system is
demonstrated here. The system first performs detection and
tracking, and uses the observed tracks to predict the motion
model probabilistically using the technique described in this
paper. The
predicted motion model is used to determine the future
locations of objects so that visibility analysis can be
performed by looking for predicted crossing between different
moving objects. As shown in the paper, time intervals during
which objects are visible can be combined with camera settings
that allow an object to be captured at task-specific
requirements, including resolution, direction of motion, etc.
The end results are a set of time periods during which a given
camera can satisfies task related to the object. These time
periods are then used for scheduling the camera using a DP
approach. (paper)
- Looking for people visible in the field of regard, even with people moving and occluding one another in a circle. This is done with constant update to the predicted motion model when the error ratio increases beyond a certain bound
- Investigating the reliability of the motion predictor. Here, the tracked individual is walking on a curved path. Red circles are the predicted bounds of the orthographic projection of the person on the ground plane in the following frames while green circles are the predicted bound in the current frame. The latter can be compared with the actual location given as a blue bounding box.
- Illustrating using the active camera system for face capture.
- Illustrating the effect of increasing task-required resolution. Here, the resolution required is lower, and two active cameras are needed to capture the robots. In the video, the resolution was increased and three cameras are now needed.
- Increased resolution requirement.
- Capture of three persons through occlusion. A CONDENSATION tracker is used to track the people through short periods of occlusions, and the observed tracks are used for capturing full-body videos of the people.
- Unattended baggage. A very robust approach that do
not use thresholding, that is the pitfalls of many vision
systems. (paper)