Angular>> facetracker>> 返回
项目作者： huangfcn

项目描述 ：
detect face using MTCNN and tracking face with KCF/Kalman Filter. The assignment problem solved by Hungarian.
高级语言： Objective-C
项目地址: git://github.com/huangfcn/facetracker.git
创建时间： 2019-11-16T05:11:33Z
项目社区：https://github.com/huangfcn/facetracker
开源协议：
下载


Description

  1. This project is very similar to my another project multiple-object-tracking, the only difference 
  2. is a general YOLO3 detector in that project replaced by the MTCNN model (detecting faces only). 
  4. Face detection and tracking is one of the most widely used application of deep-learning. 
  5. Face detection model only detects faces visible on current frame. To track obstacle faces and moving
  6. faces, we need to solve a multiple-object-tracking problem. This could be done in two steps:
  8. 1, Detecting moving faces in each frame
  9. 2, Tracking historical faces with some tracking algorithms
  11. An assignment problem is used to associate the faces detected by detectors and tracked by trackers.
  13. We can found some introduction of this framework here,
  14. https://towardsdatascience.com/computer-vision-for-tracking-8220759eee85
  16. Another example in more detail with matlab code (detecors and trackers may different),
  17. https://www.mathworks.com/help/vision/examples/motion-based-multiple-object-tracking.html
  19. Here I implemented a highly efficient and scalable C++ framework to combine the state of art 
  20. deep-learning based face detectors (MTCNN here) and correlation filters based trackers 
  21. (KCF, Kalman Filters also implemented). The assignment problem is solved by hungarian algorithm.

Detectors: MTCNN

  1. MTCNN is a specific model used to detect face. This model is implemted using the technology described 
  2. in my another project 'dnnsimd' and running purely on CPU with SIMD acceleration (auto-vectorization).

Trackers: Kalman Filter and KCF

  1. Kalman filter is fast but less accurate. KCF is accurate but much slower. 
  2. They are implemnted with exactly same interface, so we can easily switch from one to another 
  3. in the project.

Live Camera Capture: OpenCV

  1. OpenCV is used to capture live video frames and used for image preprocessing.

Misc

  1. 1, Only x64/Release environment variables are properly set
  2. 2, MTCNN supporting library (conv.c) can only be compiled by gcc to enable SIMD acceleration, 
  3.    a Visual Studio compatible conv.o is provided and compiled by gcc with '-O3 -march=native'.