Multi-person joint attitude tracking under complex scenes, the paper designed an on-line optimization network to establish the frame attitude relationship, and formed the attitude flow, in addition, designed a attitude flow non-maximum value suppression to reduce redundant attitude flow, re-establish the relationship between the time of the non-intersecting attitude. Experiment on two attitude Tracking database pose Tracking datasets
([Iqbal et al.] and [Girdhar et al., 2017]) contrast.
There are two kinds of attitude tracking methods, top-down and bottom-up. The top-down approach first detects body bbox, estimates key points, and tracks box. The bottom-up method detects the candidate each frame, establishes the space-time map, forms each person's attitude trajectory. The top-down approach is more promising, but obscured, truncated, blurred images, and so on, as shown in the following figure, which requires sharing information between frames.
* * Method Structure * *
Two-step, pose flow building and pose flow NMS
* * Basic concept * *
1. In-frame attitude distance
To measure the similarity of two gestures in the same frame, the soft match function is:
Where $c_1,c_2$ is the key point of Attitude $p_1,p_2$, the direct spatial distance of the key points is:
2. Inter-frame attitude distance using orb matching calculation
**pose Flow building**
Attitude estimation establishes the relationship of the same person between frames, and, for a certain gesture, creates a candidate-associated attitude set in subsequent frames, i.e.:
Where DC is the orb distance. For the From T to (t+t) frame, maximize the target function:
Among them, S (p) is the confidence score of posture P, which is related to box score and key point score.
**pose Flow nms**
1.Pose Flow Distance
For any two attitude flows, extract the sub-flows that intersect their time, and the distance of two attitude flows is:
where DF () is the intra-frame attitude distance
2. Attitude Flow Fusion
The attitude flow with the maximum confidence score as a reference, by using DPF (), fusing other flows close to the reference attitude flow, forming a new attitude flow, representing the attitude flow, the key points representing the attitude flow 2D coordinates and the confidence score are:
The schematic diagram of the attitude Flow NMS is:
* * Experimental results * *