Loading...

Improving Multi-pedestrian Tracking by Learning Appearance Model

Sabzmakan, Amin | 2014

674 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 46196 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiee, Hamid Reza
  7. Abstract:
  8. Object tracking is an important tool for video analysis which has been applied to public surveillance, road safety, and scene analysis. In multiobject tracking, the goal is to extract the trajectory of every target using spacial and temporal information. Since new objects can enter the tracking area, an object detector is required to detect their presence. This detector locates target objects in all frames and its (noisy) output is delivered to an associator that returns each object’s trajectory. The associator uses the motion and appearance model of the objects to discover their relation and find the trajectories. In this thesis, each target is partitioned into a number of non-overlapping patches. A general model is learned on features extracted from these patches to form a prior that is then used to generate specific models for each detected object. The associator uses the KL-distance between models learned for each object to quantify their similarity. To evaluate the proposed method, three video sequences are used from the PETs data set. The results of the proposed method show its success in comparison with the state-of-the-art algorithms. Specifically, the proposed method was able to improve the Multi-Object Tracking Accuracy (MOTA) measure by six units
  9. Keywords:
  10. Tracking ; Object Tracking ; Multiobjective Tracking ; Public Survilance ; Pedestrain Tracking

 Digital Object List