Activity Recognition

Moving object detection and tracking

In this paper, an edge segment based tracking algorithm, which is capable of identifying moving objects in image sequence, is proposed. Since segmenting objects from a sequence image are not easy, traditional object tracking algorithms fetch difficulty due to large variation of object shape, orientation, motion and size among frames. One object may consist of several parts with different motion. Additionally, objects’ motion and shape are less consistent within frames. To cope with these difficulties, our algorithm makes efficient use of edge segments based on the Canny edge map by utilizing the edge structure in the moving object region. Curvature-based features are used for moving edge registration due to its transformation invariance nature. We use the maximum curvature correspondences between two edge segments to define the 2D affine transformation that relates the two segments by solving a linear system. The edge segment registration error is also minimized. A Kalman filter based predictor is used for tracking each individual edge segments. Edge segments are clustered by using a weighted mean shift algorithm. Finally, a group motion tracker is used for tracking moving object from each cluster. Experiments show that our edge-segment based tracking algorithm can track moving objects or part of the object efficiently under varying illumination conditions and partial occlusion.

  • Edge information is extracted from video frames and represented as segments using edge class.Three edge lists are maintained:
    • Initial Reference Edge List – stores the accumulated background edge information.
    • Temporary Reference Edge List – stores the temporary changes in dynamic background.
    • Moving Edge List – stores the moving edge information.
  • Eliminates ghost effect, makes the system robust against illumination and temporary changes of background.

Motion estimation

Detecting dominant motion patterns from video is important due to the increasing demand in video surveillance problems like traffic analysis, human behavior analysis, airport security, patient monitoring, customer behavior monitoring, detecting shooting parameters for mobile devices etc.

  • Allows both Camera and Subject (multiple) motion estimation.
  • A wide range of motions are detectable.
  • Provides a rough shape and positional data of objects.

Comments are closed.