IOT Tracker

1. Introduction

최근 추적 기법 트랜드 : tracking-by-detection

1. first an object detector is applied to each video frame.
2. In a second step, a tracker is used to associate these detections to tracks.

챌린지

• limited performance of the underlying detector which may produce false positive and missed detections
• MOT
• Paths become ambiguous

기존 연구 결과 Many methods have been proposed to solve these problems:

• [1, 2] define a continuous energy function and search for strong local minima using sophisticated minimization techniques.
• [6] estimates short tracklets for unambiguous frames and stitches them according to a dynamics-based similarity.
• Other approaches include using a globally optimal and locally greedy method and integer linear programming [12] and online discriminative appearance learning [3].

[1] A. Andriyenko and K. Schindler. Multi-target tracking by continuous energy minimization. In Proceedings of the IEEE Conference on  (CVPR), pages 1265–1272. IEEE, 2011.
[2] A. Andriyenko, K. Schindler, and S. Roth. Discretecontinuous optimization for multi-target tracking. In Proceedings of the IEEE Conference on  (CVPR), pages 1926–1933. IEEE, 2012.
[6] C. Dicle, O. I. Camps, and M. Sznaier. The way they move: Tracking multiple targets with similar appearance. In Proceedings of the IEEE Conference on  (CVPR), pages 2304–2311, 2013.
[12] H. Pirsiavash, D. Ramanan, and C. C. Fowlkes. Globallyoptimal greedy algorithms for tracking a variable number of objects. In Proceedings of the IEEE Conference on n (CVPR), pages 1201–1208. IEEE, 2011.
[3] S.-H. Bae and K.-J. Yoon. Robust online multi-object tracking based on tracklet confidence and online discriminative appearance learning. In Proceedings of the IEEE Conference on CVPR,pages 1218–1225, 2014.

본 논문의 기본 아이디어 : in this paper a very simple tracking approach shall be assessed which is based on the idea of a passive detection filter introduced in [8].

[8] V. Eiselein, E. Bochinski, and T. Sikora. Assessing post-detection filters for a generic pedestrian detector in a tracking-by-detection scheme. In Analysis of video and audio ”in the Wild” workshop at IEEE AVSS17, Lecce, Italy, Aug. 2017.

2. Method

가정사항

• The detector produces a detection per frame for every object to be tracked,
  • i.e. there are none or only few ”gaps” in the detections.
• detections of an object in consecutive frames have an unmistakably high overlap IOU

정의 : We propose a simple IOU tracker

• which essentially continues a track by associating the detection with the highest IOU to the last detection in the previous frame if a certain threshold σIOU is met.
• All detections not assigned to an existing track will start a new one.
• All tracks without an assigned detection will end.

• $$D_f$$ denotes the detections at frame $$f$$
• $$d_j$$ the $$j^{th}$$ detection at that frame
• $$T_a$$ active tracks
• $$T_f$$ finished tracks
• $$F$$ the number of Frames in the sequence

Note that in line 5 only the best-matching, unassigned detection is taken as a candidate to extend the track.

This does not necessarily lead to an optimal association between the detections Df and tracks $$T_a$$ but could be solved

• e.g. by applying the Hungarian algorithm maximizing the sum of all IOUs at that frame.

However, taking the best match is a reasonable heuristic since $$σIOU$$ is normally chosen in the same range as the $$IOU$$ threshold for the non-maxima suppression of the detector.

Therefore, multiple matches satisfying $$σIOU$$ are rare in practice.

3. Experiments

4. Conclusions

Our presented IOU tracker considerably outperforms the state-of-the-art at only a fraction of the complexity and computational cost.

This becomes possible due to the recent advances in the object detection domain, not at least due to the current boom of CNN-based approaches.

Tracking Things in Object Detection Videos의 정리글

동작 방식

It's incredibly simple, it works by comparing the overlapping areas between the two detections between the frames.

It does this by computing the intersection over union or the areas: $$ IOU(a,b) = \frac{Area(a) \cap Area(b)}{Area(a) \cup Area(b)} $$

And it does just that (no prediction, no velocity vector computation...)

제약 Limitations

• 예측을 하지 않아 특정 프레임에서 탐지를 하지 못한다면, 다음 프레임에서 새로운 ID를 할당 Doesn't do any prediction, so if YOLO loses the object for some frames, the tracker will lose it as well and will track it again under a new id.

• Won't perform well on lower frame rate detections, this is understandable, the overlapping areas at lower frame rates can be non existent as there is no predictions.

How does it perform on our use case of tracking cars ?

• Surprisingly great !

• We think out of the box it may not be as good as our current algorithm tracker for some case because YOLO is missing detections quite a lot of times and triggers lots of re-assignments with this tracker...

• but it had a huge potential of improvement if we add prediction + re-entering features.

Takeaways from the IOU tracker

SIMPLE IS THE BEST : “simple tracking methods like the IOU tracker can lead to better results than complex approaches based on decades of research”

기존 유클리드 기반 distance()함수를 IOU로 바꾸면 성능 향상 가능 Based on what we learned we revisited ourdistance()function. It could be improved by using this overlapping area comparison.