高速跟踪--High-Speed Tracking-by-Detection Without Using Image Information

高速跟踪: 当检测精度较高，视频帧率较高时，跟踪问题就会变得很简单，主要是多阈值目标检测和 判断前后帧的重合率
High-Speed Tracking-by-Detection Without Using Image Information
Advanced Video and Signal Based Surveillance (AVSS), 2017 14th IEEE International Conference on

https://github.com/bochinski/iou-tracker/

both high precision detections and the usage of video footage with high frame rates can greatly simplify the tracking task
当检测的精度很高，视频的帧率也很高时，跟踪就会变成一个很简单的问题。
If both requirements are met, tracking becomes trivial and can be done even without using image information

The IOU measure used in our approach is defined as

这里基本的思路可以看作是借鉴 Canny 算法中的 双阈值，高阈值得到确定的检测框，低阈值包含所有可能的检测框，使用前后帧信息将漏检的物体框识别出来加入到确定的检测框中去。再根据跟踪持续的时间及跟踪中检测的置信度来过滤

这种跟踪基本不耗时

当跟踪变形较大的物体时，这种简单的跟踪算法效果就会稍差些

in the case of vehicle tracking and dealing with fixed-sized objects, static cameras and high accuracy detections at high frame rates, good tracking can be achieved on a simple level.

相关文献还有一篇：
Assessing Post-Detection Filters for a Generic Pedestrian Detector in a Tracking-By-Detection Scheme
这里提出了一个 Passive Detection Filtering Using Temporal Hysteresis Thresholding 和 Active Detection Filtering Using Optical Flow
Passive Detection Filtering ：基于单帧的检测使用双阈值高阈值得到确定的检测框，低阈值包含所有可能的检测框
Active Detection Filtering ：基于光流信息得到运动物体，弥补单帧目标检测的一些缺陷

11