| 摘要 |
For tracking objects in a sequence of video frames, an object detector (11) detects objects, and a data association (112) is established between an object tracker (121) and a detection (110) by evaluating online (on-the-fly) trained instance-specific classifiers for each object For tracking (52) the detected objects (110), particle filtering is employed by determining weights of importance for the particles associated with an object tracker, and by propagating the particles using a motion model. For the positions of the particle associated with an object tracker, the detector confidence (111) is determined from the object detector, and the weights of importance are determined for the particles based on the detector confidence at the positions of the particles. By exploiting the intermediate, continuous detector confidence for tracking purposes, tracking is not relying solely on the detector's final and discrete indication of a detection, and the flexibility for handling difficult situations is improved. |
