A fixed video camera records people moving through its field of view.
The goal is to detect and track the head, in real-time as it moves through the video. The norm is there are many heads, which often are sometimes partially obscured. This example video boxes heads and provides a head count.
There seems to be many different models. Examples include:
Given the context of the video, what is the thought process that you would use to choose a model?
There may be a conceptual disconnect between the term Semantic Detection and the task of Head Tracking, since sequential recognition of an object in a set of visual samples representing continuous movement isn't technically a semantics problem. Although a mapping strategy that works with semantic processing may somehow apply to, with appropriate modifications, the mapping of obstructions and tracked objects and their relationship in space-time.
Tracking a head is primarily a surface fitting convergence problem, where the convergence goal and the process that achieved it must be sustainable, especially if the system is required to take actions during the tracking operation. Distinguishing heads from one another requires a high degree of reliability and accuracy in this convergent continuity. That is the primary challenge.
Drawing boxes and displaying counts are trivial operations once the head tracking works.
These are some models listed in the question or improvements over them.
One seasoned expert spoke to a particular detail in the final question.
The thought process experts use to choose a model has no equivalent point value in StackExchange reputation. As experts retire, some may be willing to sell the professional heuristics they acquired over decades of research and development work if they haven't done so already to a book publisher.
Nonetheless, the above links are a good furtherance of the thought process begun in this question along a direction of continued effective thought. The following steps provide a road map for how to develop approaches in general.
