It can be said that the most critical task of multi-object tracking (MOT) based on deep learning is not to recognize the target, but to detect the target 遮挡后重识别
. There are a large number of trackers available, but not all of them have good re-identification pipelines. In this blog post, we focus on one such tracker FairMOT
that revolutionizes the joint optimization of detection and re-identification tasks in tracking.
FairMOT-Result
1. Review of multi-object tracking
1.1 Types of trackers
Trackers can be classified according to some attributes
- (1) Number of objects tracked
- Single object trackers, including traditional OpenCV trackers such as CSRT, KCF, etc. They work by initializing an object in the first frame and tracking it throughout the sequence.
- Multiple object trackers include deep learning trackers that are trained on a dataset to track multiple objects of the same or different classes. These include DeepSort, JDE, FairMOT, and more.
- (2) Detection pipeline
- Traditionally, box coordinates are manually initialized around the object in the first frame. These objects are then tracked in the next frame. These are trackers without an instrumentation pipeline.
- Recent algorithms have an object detection pipeline built together with an association stage for better tracking results. These are trackers with detection pipelines.