Hello. A lot of work has been done on this problem. Research started in the 1960s or thereabouts in radar, then in sonar and video. Most of the advanced techniques have been applied in radar systems and have gradually percolated into the latter two areas, although I think JPDA (joint probabilistic data association) might have been done in sonar first. In video tracking, you usually track image centroids whereas tracking extended objects is a more recent development.
The list of approaches is long and was long even before the appearance of more recent techniques such as particle filters and methods based on finite set statistics (PHD, etc), which ignore target labels. The simplest is the nearest neighbour Kalman or alpha-beta filter. The most complicated is MHT (multihypothesis tracking). Particle filters can be used as can Gaussian mixture filters, but I would only use them if the traditional techniques failed (or I wanted to be more fashionable :-) Watch out for comparisons between apples and oranges (e.g. recursive versus batch).
Classical approaches are summarised in the paper linked to below. There is no reason that any of these could not be applied to video tracking based on sequences of 2-D images.
Article Taxonomy of multiple target tracking methods
Hello Eren, Here i have added few point which are important while working with multi objects with static camera videos, so please go through these points.
Object tracking itself is the task of following one or more objects in a scene, from their first appearance to their exit. An object may be anything of interest within the scene that can be detected, and depends on the requirements of the application. Given a sequence of image frames to trace a set of objects, which are sub images, in each frame. In general, in a dynamic environment both background and object are allowed to vary. In principle, to solve this general unconstrained problem is hard. One can put a set of constraints to make this problem solvable. The more the constraints, the problem is easier to solve. Some of the constraints that generally imposed during object tracking are:
1. Object motion is smooth with no abrupt changes
2. No sudden changes in the background
3. Gradual changes in the appearance of object
4. Fixed camera
5. Number and size of objects
6. Limited amount of occlusion
Most of the methods I know are based on some kind of particle filter or similar approach. If you're trying to track a specific class of objects (pedestrians, cars, ...) there are some papers on tracking by detection/recognition using HoG/SVM, Hough forests etc.
Best performance is usually achieved with so-called offline optimization methods, where an objective function over a certain time span (or the entire sequence) is minimized. Examples of such methods include:
LP (ILP) and k-shortest paths. Works best with multiple static cameras
http://cvlab.epfl.ch/software/ksp
Continuous / Discrete-continuous energy minimization
http://research.milanton.net/contracking/
http://research.milanton.net/dctracking/
Hierarchical / energy-based tracklet association
http://iris.usc.edu/people/yangbo/research.html
There is also some work, specifically concerning team sports, e.g.:
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6619083
Hello. A lot of work has been done on this problem. Research started in the 1960s or thereabouts in radar, then in sonar and video. Most of the advanced techniques have been applied in radar systems and have gradually percolated into the latter two areas, although I think JPDA (joint probabilistic data association) might have been done in sonar first. In video tracking, you usually track image centroids whereas tracking extended objects is a more recent development.
The list of approaches is long and was long even before the appearance of more recent techniques such as particle filters and methods based on finite set statistics (PHD, etc), which ignore target labels. The simplest is the nearest neighbour Kalman or alpha-beta filter. The most complicated is MHT (multihypothesis tracking). Particle filters can be used as can Gaussian mixture filters, but I would only use them if the traditional techniques failed (or I wanted to be more fashionable :-) Watch out for comparisons between apples and oranges (e.g. recursive versus batch).
Classical approaches are summarised in the paper linked to below. There is no reason that any of these could not be applied to video tracking based on sequences of 2-D images.
Article Taxonomy of multiple target tracking methods
As previously commented, there are a lot of approaches presented in the last decades. You can see a very good review of the more recent state of the art in our last publication, where a new tracking method for multi objects with static camera videos is presented:
http://dx.doi.org/10.1016/j.eswa.2015.06.016
Good luck with your research!
Article Expert Video-Surveillance System for Real-Time Detection of ...
- Badges
- Science topic