当前位置：网站首页>One of IOU target tracking: IOU tracker

One of IOU target tracking: IOU tracker

2022-07-27 21:10:00 【Lingshan】

IOU MOT firstly ：IOU Tracker

1. Introduce
2. Algorithm content

Reference:

High-Speed Tracking-by-Detection Without Using Image Information

1. Introduce

IOU Tracker yes 2017 A simple and efficient tracking model without image information was proposed in , It's in DETRAC Vehicle tracking data sets can be easily run to 100K fps.

It is mainly based on two assumptions ：

The detector generates a detection for each object to be tracked in each frame , in other words , In the process of testing No, / Only a few Discontinuity of ;
Detect an object with high overlap in consecutive frames $I O U$ , This is common at a sufficiently high frame rate .

Used in the text $I O U$ Measurement method ：
$\operatorname{IOU}(a, b)=\frac{\operatorname{Area}(a) \bigcap \operatorname{Area}(b)}{\text { Area }(a) \bigcup \operatorname{Area}(b)} .$

If both conditions are met , Tracking will become less important , Even without using image information .

2. Algorithm content

What is put forward in this paper $I O U$ tracker , If a certain threshold is met $σ_{IOU}$ , It passes the highest $I O U$ Associate with the last detection of the previous frame to continue tracking . All probes that are not assigned to an existing trace will start a new trace . All traces of tests that are not assigned will be killed .
Insert picture description here By filtering out all lengths less than $t_{min}$ I didn't give a score for the tracking and once $σ_h$ The above tests , Further improved performance . Shorter traces are deleted , Because they usually come from false positives , It will generally increase the confusion of output . At least one high score test is required for tracking , To ensure that the trace is an object of real interest .

The detailed method description is shown in the figure below , among $D_f$ Said in the first $f$ Frame detection , $d_j$ It means at the... Of that frame $j$ Test results , $T_a$ Active (active) track , $T_j$ Completed tracking , $F$ The number of frames in the sequence ：
Insert picture description here
technological process ：

For the current frame , First, use the threshold $\sigma_l$ Make a preliminary screening , Get the input detection set $D$ ;
For each active trace , Find and it has the largest $I O U$ The frame detection box , If it's time to $I O U$ Greater than $\sigma_{IOU}$ , Will detect $d_{best}$ Join the currently active tracking $t_i$ in , Then, the detection frame is detected from the current frame set $D_f$ Delete inside ; And if the biggest $I O U$ Less than $\sigma_{IOU}$ , At this time, if the historical highest detection score of the active tracking is greater than $\sigma_h$ And the tracking time is longer than $t_{min}$ , Will $t_{i}$ Add to $T_f$ in , Otherwise it will be from $T_a$ Delete from $T_a$ ;
For detection that does not match , Initialize it as a new trace and add it to $T_a$ in ;
After the above cycle , Track activation $T_A$ Make judgments for each trace in , If the highest detection score in the history of tracking is greater than $\sigma_h$ And the tracking time is longer than $t_{min}$ , Then the trace $t_i$ Add to $T_f$ in .