IOU target tracking II: viou tracker

Reference：

1. Extending IOU Based Multi-Object Tracking by Visual Information

1. Introduce

In the last version of IOU Tracker in , An extremely simple but accurate tracking method is used , Rely entirely on input detection as a high speed IOU tracker. For real-world applications , Highly dependent on potential detectors for high recall , Gaps caused by one or several missed tests will not only lead to false negatives (FN), It will also cause the termination and restart of the trace , This leads to a high fragmentation rate and ID Switching rate ------ This greatly reduces the quality of tracking .

In this article , The question is passed When no target is detected , Using vision to track a single target to improve the robustness against detection loss to overcome .

2. Algorithm content

Visual tracking can be carried out in two directions : First of all , If the tracking does not meet the above IOU Tracker Medium ${\sigma }_{IOU}$ threshold , At the last known target location ( That is, the detection of the previous frame ) Initialize the visual tracker on , And the longest goal $ttl$ Frame tracking ; If new tests are in these $ttl$ Intra frame satisfaction ${\sigma }_{IOU}$ threshold , Then stop visual tracking , continue IOU track . conversely , Trace stop .

However , As the number of visual tracking frames increases , It becomes more likely that the visual tracker loses track or jumps to another object . In order to limit the number of consecutive frames that only visual cues track the target , For each new trace, the last... Is also performed to the previous frame $ttl$ Visual tracking of frames ( Reverse tracking ). If the overlap condition satisfies an existing , The tracking has ended , We merge them . In this way , It can be closed with a length of no more than $2\cdot ttl$ Frame gap , A single visual target tracker is only used for maximum $ttl$ frame .

Although visual forward and backward tracking of objects helps to merge interrupted tracking tracks , But it also adds visual tracking endings at the beginning and end of each completed track , As shown in the middle of the above figure . Because tracking should start when the object enters the scene , End when the object leaves , These visual tracking terminals do not promote merging gaps , And easy to contain errors , Because the object of interest may not be in the scene . For this reason , This paper limits the use of visual tracking , And remove the visual tracking terminals of these visual tracking bounding boxes from the tracking .

About visual tracker The selection of , Used in the paper KCF(2015) and Medianflow(2010) detector .

Article jump ：
IOU Target tracking is one of them ：IOU Tracker