Extending IOU Based Multi-Object Tracking by Visual Information论文简读

EDITORIAL

This paper is in the process of "High-speed tracking-by-detection without using image information" (IoU) based on the improved version, mainly to solve the former in the case of loss detection, and identity missed the conversion problem prone.

contribution

1. A visual tracking using a single pair IoU tracker be extended to predict the position of the track at the time of detection of the next frame is lost.
2. We proposed a batch of tracking mode, effectively reducing the IoU tracker identity conversion tracking and debris, thereby enhancing the MOTA value IoU tracker.

Tracking Process

IoU tracking process

We first review the tracking process IoU tracker:

1. For the current frame, using the first threshold value $\sigma_l$Detection was filtered to obtain an input current detection $D$。
2. For each active tracking $t_i$, The largest find of its final position detection IoU $d_{best}$If satisfied $IOU (d_ {best}, t_i) \ geq \ sigma_ {IOU}$, The latest location tracking updates $d_{best}$,will $d_{best}$From the detected set $D$ removed. If not $IOU (d_ {best}, t_i) \ geq \ sigma_ {IOU}$, It is determined by detecting the highest scoring track corresponding to the detected value is greater than a threshold $\sigma_h$And is greater than the length of the track $t_ {min}$To measure track $t_i$Whether it is a complete tracking. in case $t_i$Is a complete tracking, add it to the full track list $T_f$If not then it is terminated. The track from activating the tracking $T_a$Removed.
3. For the detection of not matched, it is initialized to a new track and added to the active trace $T_a$in.
4. After completion of all frames repeating the steps above, it is determined by activating the tracking $T_a$Each trace corresponds to the maximum detection points detected is greater than a threshold value $\sigma_h$And is greater than the length of the track $t_ {min}$,to measure $t_i$Whether it is a complete tracking. If so, add it to the track to complete $T_f$in.

V-IoU extension

V-IoU trackers in the tracking process IOU tracking algorithm, adding a single visual target tracker. Add it in supplement the main track process:

1. In the step 2, not matched for tracking the use of a single visual target tracker prediction. If successful prediction, using the predicted position in the tracking position as the current frame, the tracking track added to the active, if the prediction fails, it is determined whether a complete track. In the absence of detection matching case, such a single visual target tracking process will continue maximum $ttl$ frame period and if a successful match is detected, a single visual object tracking is terminated by detecting a location update tracking position, until the next detection of a match failure.
2. After detecting the new track is initialized to reverse its maximum $ttl$ visual tracking a single target frame, if the tracing and tracking the completion of the previous time and space to meet the standards overlap, then the two merged track.
3. For the merger of two tracks, remove those overlapping visual tracking and tracing distant.

Personal feelings

V-IoU tracker, also followed the tracking-by-detection paradigm, the biggest feature is the use of a single target tracker to track the predicted no match detection and tracking, which is the future trend of multi-target tracking, after the paper Some multi-target tracking algorithm, are also followed this paradigm.