Hello :),
I have been using OC-SORT in my project and I wanted to understand it on a deeper level. Specifically, I'm trying to understand how the Observation-based Re-association Update (ORU) mechanism works as illustrated in Figure 7 of the paper. The visual representation shows the tracking process before and after an object becomes untracked, but I'm unclear about exactly how ORU generates the updated state estimation (shown in red) after re-association at zt₂.
Specifically, I can see how the tracking works before the object becomes untracked, but I'm confused about the transition at zt₂. The first two red estimations appear to be linear extrapolations of past observations, but the final ORU estimation after zt₂ follows a different trajectory. How does ORU determine the correct velocity direction after re-association?
My current understanding is as follows:
In figure 7b, we observe that a detection after zt2 (blue triangle) has not been associated. This is due to the KF not being able to get a correct prediction for an object's position. Now, since we have the re-association at time zt₂, we can use the track history for the object to inform a velocity direction vector using this zt₂ and zt1. Then we can use this to inform the motion direction more aggressively than the Kalman filter would normally allow. Hence, we are able to obtain re-associations in the next step. Is this understanding correct?
I'd appreciate any help on this.
Thanks!
Hello :),
I have been using OC-SORT in my project and I wanted to understand it on a deeper level. Specifically, I'm trying to understand how the Observation-based Re-association Update (ORU) mechanism works as illustrated in Figure 7 of the paper. The visual representation shows the tracking process before and after an object becomes untracked, but I'm unclear about exactly how ORU generates the updated state estimation (shown in red) after re-association at zt₂.
Specifically, I can see how the tracking works before the object becomes untracked, but I'm confused about the transition at zt₂. The first two red estimations appear to be linear extrapolations of past observations, but the final ORU estimation after zt₂ follows a different trajectory. How does ORU determine the correct velocity direction after re-association?
My current understanding is as follows:
In figure 7b, we observe that a detection after zt2 (blue triangle) has not been associated. This is due to the KF not being able to get a correct prediction for an object's position. Now, since we have the re-association at time zt₂, we can use the track history for the object to inform a velocity direction vector using this zt₂ and zt1. Then we can use this to inform the motion direction more aggressively than the Kalman filter would normally allow. Hence, we are able to obtain re-associations in the next step. Is this understanding correct?
I'd appreciate any help on this.
Thanks!