It is possible, I did once for multi object!

But I believe the multi-tracker is much better than multi single-tracker.

For the sake of clarity, assume a single tracker as an agent which is responsible to track a person.  Now for multi tracker you have a multi agent system. 

In general, in multi agent systems, cooperation between agents yields a more reliable solution than applying many agents acting individually. In your specific application robustness could be followed by handling correctly collision between two trackers, for instance. One solution might be having a centralized system that monitors all the agents' activities and in such a critical events (aforementioned example) makes a decision according to the history and estimation of all the agents.

Hope it gives you a general idea.

It is possible but performance decreases proportionally to the number of targets because you will have a single-tracker for each one. One I approach I would consider is the Probability Hypothesis Density (PHD) filter, which is based on Random Finite Set (RFS) theory. It performs multitracking of all target of the scene estimating the next state of all current targets at once, but it lacks identification of the targets.

It is posible, you make copies of your  single object tracking routine and make then act as independent agents. Next you have a define a "loose control" method which enables agents to work in an independent but coordidinate way.  

Hello again,

This is one of those questions that will separate academic researchers from practitioners. Most tracking algorithms used in practical systems are single target except for very high performance radar applications (ATM and GBAD). The majority of tracking algorithms are pretty poor when it comes to multiple targets and single target methods tend to work better (!). The exceptions to this are things like MHT, which is hard to implement, and certain assignment algorithm approaches (>2-D), which are computationally intensive. The "modern" approaches, like PHD implemented with Gaussian mixtures, have not been proven to outperform conventional approaches like MHT, and at any rate, they are both basically propagating multiple Gaussian hypotheses. I think that engineers should be scientific in their approach and actually compare and benchmark their methods properly. In MTT this is not the case however. 50+ approaches and 10,000s of papers and still no one knows what works best on a given problem or can offer any useful design rules. This is not scientific.

By the way, single-target VDA, which is a track splitter, performs really well in MTT despite. Have a look at my RG page for references, starting with the attached link. There is a paper by Musicki that takes a similar approach.

Another thing about MTT approaches is they often assume a constant number of targets (like JPDA). MHT and PHD, which don't do this, use pruning. You can also use clustering with splitting and merging, as in the second reference.

Article Taxonomy of multiple target tracking methods

Patent Method for Cluster Management in Multiple Target Tracking