Particle filtering, PSO, mean shift, Kalman filter are used for tracking objects within video sequences. Under your consideration which one is the most accurate one? Do you have in mind any code available?
I suggest you to read this paper, the codes are available online at the link below.
Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang, “Online Object Tracking: A Benchmark,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013.
Mean-Shift algorithm is better for tracking when preceded by a detection algorithm. It have low computational cost and is robust. The other option is Scale Invariant Feature Transform (SIFT) which is almost invariant to many changes in object, like shape, rotation and translation etc.
I have worked on Mean-shift implemented it in real-time. however you can also find the source code of both the techniques on codeforge.com.
Geia sou Kostantine, have a look at https://www.researchgate.net/publication/259621328_Fast_Object_Re-detection_and_Localization_in_Video_for_Spatio-Temporal_Fragment_Creation and let me know if you need any further help on this.
Conference Paper Fast Object Re-detection and Localization in Video for Spati...
Hello. At the present time in the target tracking community (including for video sequences) it is not possible to say what the "best" object tracking algorithm is, even if you define what you mean by "best", whether this be most accurate, best sensitivity, etc.
In the last 10 years or so thousands of papers have been written on particle filtering, PHD filters, etc. Prior to this, at least for recursive tracking as opposed to batch, the state of the art was considered to be MHT (multihypotehsis tracking).
AFAIK There is unfortunately only 1 paper that compares MHT and the PHD filter and no papers comparing particle filtering with MHT. The one comparison that exists in the public domain is not conclusive and used a home-grown MHT, not a commercial grade one. Another important observation is that almost all tracking papers use a 2-D measurement and 4-D state. This means that any false alarms are distributed in a 2-D volume that is typically visualised by plotting all the measurement data on a 2-D plot with time on the x-axis. This gives the impression of a huge clutter density when the actual clutter density is in fact very low (compared, for instance, with passive sonar data).
I hope people find these comments thought provoking, since they are not aligned with popular thinking on the subject. For further information, please refer to the survey paper below.
If you wish to fill the gap and perform a comparative simulation, please feel free to use the benchmark sonar data sets described below and available from ResearchGate. These data sets are for 1-D measurements and are very challenging - i.e. a basic tracker like the nearest neighbour Kalman filter often fails on this data.
See the third link for further comments on the false alarm density issue. This paper compares a MHT-like approach with several other conventional approaches in terms of track loss and track swap performance - which are best judged separately, rather than using a combined or all-in-one metric.
Disclaimer: All comments are my own and do not reflect the views of my employer.
Article Taxonomy of multiple target tracking methods
We have recently published a new algorithm for people tracking which is very effective and efficient. You can see this in the results obtained in our paper, which are compared to the main state-of-the-art methods:
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 ...
It depends what application type u address, if you need robustness particle filtering is a good one, Motion-based techniques like background subtraction are appropriate for real-time applications
I can answer on probabilistic methods, though there is no definitive quantification of BEST.
1. For linear measurement and motion model, Kalman filter is the best to minimize mean square error or to maximize measurement likelihood.
2. In case of high non-linearity, particle filter is the best in terms of accuracy, if computation time is not of concern. UKF is a good compromise between computation and accuracy. EKF is best in terms of computation time.
3. For angle-only measurements, shifted Rayleigh filter is the best as it almost matches particle filter for most scenarios.
4. If we ignore particle filter, for radar sensors, Decorrelated un-biased converted measurement KF is one of the best.
5. For distributed tracking without the knowledge of cross-correlation between estimates, Information Matrix fusion is optimal.