Tracking Benchmark
2013年Visual Tracker Benchmark对多种算法进行了一些评测,本文主要是在它的基础上进行了一些总结。
说明
1. 下面的分析基于单个物体跟踪
2. 被跟踪物体的大小对性能的影响没有考虑进来
Mean Shift
Year | Name | Model | Model Update | Search Mode | Adaptive Size | code | FPS | Size | Pros | Cons |
1998 | Camshift[6] | - | No | - | No | C++ | ~300 | - | 1.速度快 2.原理复杂度低 |
1.结果依赖于直方图反射的结果,效果较差。如在人脸跟踪中,其会跟踪到所有肤色 2.对噪声比较敏感 |
2003 | KMS[5] | - | No | - | No | C++ | ~3000 | - | 1.速度快 2.原理复杂度低 |
Tracking by detection
Year | Name | Model | Model Update | Search Mode | Adaptive Size | code | FPS | Size | Pros | Cons |
2015 | KCF[1] | - | Yes | - | No | Matlab | 200+ | - | 1.基于CSK改善了被遮挡时跟踪失败的问题 2.效果与StrucK和TLD可比较 |
1.目标消失以后,无法找回来 |
2012 | CSK[2] | - | Yes | - | No | Matlab | 250+ | - | 1.速度快 2.原理复杂度低 3.抗噪声 |
1.被遮挡时,容易导致跟踪失败 2.目标消失以后,无法找回来 |
2011 | Struck[3] | - | Yes | - | No | C++ | 20.2 | - | - | 1.速度慢 2.原理复杂度高 |
2010 | TLD[4] | - | Yes | - | No | C++ | 28.1 | - | - | 1.速度慢 2.原理复杂度高 |
References
- High-Speed Tracking with Kernelized Correlation Filters
- Exploiting the Circulant Structure of Tracking-by-detection with Kernels
- Struck: Structured Output Tracking with Kernels
- P-N Learning: Bootstrapping Binary Classifiers by Structural Constraints
- D. Comaniciu, V. Ramesh, and P. Meer. Kernel-Based Object Tracking
- Computer Vision Face Tracking For Use in a Perceptual User Interface