Object Detection in Images/Video
Object detection is a very important step for image/video understanding. This is generally very difficult when the target exhibits large appearance variations. Although the recent years have witnessed a great success on face detection, those methods that are successful for the face might not be directly applicable to detecting other objects. One such example is the detection of the human (e.g., a pedestrian) that exhibits a very large uncertainty in shape deformation as well as the visual appearances.
We have studied a statistical model for detecting objects such as pedestrians. Our method employs a Boltzmann distribution to capture the prior of local deformation. It has an excellent performance for detecting pedestrians under partially occlusions and in clutters. A more appealing feature of this method is its parallelism for potential hardware implementation.
Component-based detectors have demonstrated their promise by breaking down the large uncertainties. However, a critical issue in handling the imperfectness of part detections is not well addressed. We have recently performed an important theoretical analysis on the tolerance ability of missing parts, and designed a general method detector ensemble that demonstrates an excellence performance in experiments.
In addition, video provide more clues than a single image for detecting those targets whose motion has special patterns. Such motion features can be very powerful for efficient object detection and identification. We have also studied motion segmentation, a very basic vision problem, hoping to alleviate the difficulties in object detection.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|