A survey of multi-objective tracking (III.)

Overview of multi-objective tracking (iii) multi-target tracking Multi-Object Trackingcodesdata Set

Multi-target tracking common data sets:

1477969643722.jpg

Ways to expose your code:

1477969661474.jpg

Or the common data set is better, so that you can directly compare the results of other people's experiments, do not repeat the experiment.

Future directions

2. MOT with video adaptation. Most detection-based MOT methods, which often apply to scopes that rely on offline learning to obtain a detector, have found a more general purpose, that is, the Mot method for any data is a research hotspot.

4. Balance between crowd density and completeness of objection representation. Often in the dense environment, the target is very small, so that the traditional representation of the apparent method is weak, how to find a small target object still has a strong ability to identify the expression is still a research direction.

6. MOT under multiple cameras. Multi-view data can restore stereoscopic vision, more in line with biological vision, I prefer this direction, must be able to dig something.

8. Multiple 3D object tracking. 3D target tracking, this is not the same as 3, multi-vision can build 3D environment, but can not build 3D target

10. MOT with scene understanding. The target tracking problem, such as the semantic information of the scene, is combined with the scene comprehension.

12. MOT with other computer vision tasks. Multiple visual tasks are combined to complement each other. such as gait recognition, posture recognition and target tracking together to deal with, information sharing, the effect may also be good.

Personally think, human for multi-target tracking processing is very good, not to mention the machine over the human, if you can ensure that the machine has the same processing mechanism with human, by virtue of the accuracy of its operations, the results will be better than human. So I prefer both binocular vision and the combination of the scene to understand the Mot method, in addition, human beings in the processing of multiple target tracking is not at the same time, should be similar to the way of time slices, I think it is worth learning from.

A survey of multi-objective tracking (III.)