Tracking system with fused motion and object detection
A detection system fusing motion detection and object detection for various applications such as tracking, identification, and so forth. In an application framework, model information may be developed and used to reduce the false alarm rate. With a background model, motion likelihood for each pixel, for instance, of a surveillance image, may be acquired. With a target model, object likelihood for each pixel of the image may also be acquired. By joining these two likelihood distributions, detection accuracy may be significantly improved over the use of just one likelihood distribution for detection in applications such as tracking.
1. A detection method comprising:
streaming video with a camera;
detecting motion in the video step for providing a motion likelihood image;
detecting an object in the video step for providing a model likelihood image;
fusion of the motion likelihood image and the model likelihood image;
wherein detecting object in a training phase comprises:
providing positive image samples of an object;
an extraction step for detecting a plurality of interest points on an object and extracting local image patches around the interest points;
agglomerative clustering step for generating a codebook;
wherein detecting an object in the video comprises:
extracting a plurality of image patches around each of the plurality of interest points on the object;
clustering the plurality of image patches by similarity by agglomerative clustering;
the plurality of image patches is matched with a codeword in a codebook;
wherein the fusion comprises:
computing savings for a hypothesis of the motion likelihood image and the model likelihood image;
computing savings for overlapping hypotheses of the motion likelihood image and the model likelihood image;
fusing the hypotheses to maximize a total saving of the overlapping hypotheses.
2. The method of claim 1 , further comprising tracking of the hypotheses.
3. The method of claim 1 , wherein:
at least one image comprising a plurality of pixels is provided by streaming video;
the motion likelihood image is for providing a likelihood of each pixel belonging to a foreground of the image; and
the model likelihood image is for providing a likelihood of each pixel belonging to a target in the image.
4. The method of claim 3 , wherein the fusion comprises merging a likelihood of each pixel belonging to a foreground of the image and a likelihood of each pixel belonging to a target in the image.
5. The method of claim 4 , further wherein if there is no likelihood of each pixel belonging to a foreground of the image or there is no likelihood of each pixel belonging to a target in the image, then a result of the fusion would comprise the likelihood of each pixel belonging to a target in the image or the likelihood of each pixel belonging to a foreground of the image, respectively.
6. The method of claim 5 , further comprising tracking of the hypotheses.