Face Tracker

基于mtcnn人脸检测+人脸跟踪（光流跟踪），就是要快快快稳稳稳

开发环境

ncnn Windows Vs2015 Win64

开源框架

ncnn
opencv

测试结果

引用

HyperFT：参考部分代码模块

ZQCNN:包含人脸识别系统常用组件模型库，并带有加速；

ncnn_example：包含部分模型转换文件

编译步骤：

1、修改CMakeList.txt中的opencv/ncnn路径，修改成你自己的路径；
2、mkdir build
3、cd build
4、cmake .. -G "Visual Studio 14 2015 Win64"
5、make -j4

注意事项

1：运行时依赖的动态库，需要拷贝到可执行文件同级目录，或是参考如下目录添加环境变量：https://blog.csdn.net/guyuealian/article/details/79412644

2: 跟踪测试demo中提供同步和异步跟踪方案，同步是指检测放入主线程，有检出后进行跟踪匹配；异步是指，检测模块有单独子线程，会将检出结果同步给跟踪主线程，用于匹配跟踪；

3: 在debug模式和release模式中,跟踪效果会有差异,由于debug模式中检测结果更新较慢会影响跟踪结果,建议使用release模式测试

Todo：

跟踪模块完善；3/11/2019
跟踪模块优化；24/11/2019
匹配逻辑需要修改, 当前的匹配逻辑仅仅是单纯的iou计算；
光流法的参数需要调整；
匈牙利匹配已添加, 判定参数仍需调整12/15
后序需要处理kalman相关逻辑问题;
[] 代码将进行重构,匈牙利匹配参数仍需调整
[]考虑加入模板匹配用于跟踪

Face Tracker

Based on mtcnn face detection + face tracking (optical flow tracking), you need to be fast and stable

Development Environment

ncnn Windows Vs2015 Win64

Open source framework

[ncnn] (https://github.com/Tencent/ncnn)
[opencv] (https://github.com/opencv/opencv)

Test Results

! [test] (https://github.com/sunbinbin1991/faceTracker/blob/master/data/IU_mark.jpg?raw=true)

Quote

[HyperFT] (https://github.com/zeusees/HyperFT): Reference part of the code module

[ZQCNN] (https://github.com/zuoqing1988/ZQCNN): Contains a common component model library for face recognition systems, with acceleration;

[ncnn_example] (https://github.com/MirrorYuChen/ncnn_example): Contains some model conversion files

Compilation steps:

1. Modify the opencv / ncnn path in CMakeList.txt to your own path;
2, mkdir build
3, cd build
4, cmake .. -G "Visual Studio 14 2015 Win64"
5, make -j4

Precautions

1: The dynamic library that the runtime depends on needs to be copied to the same directory as the executable file, or refer to the following directory to add environment variables: https://blog.csdn.net/guyuealian/article/details/79412644

2: The tracking test demo provides synchronous and asynchronous tracking solutions. Synchronization means that the detection is placed in the main thread and the tracking is matched after detection. Asynchronous means that the detection module has a separate sub-thread and will synchronize the detection result to the tracking master Threads for matching tracking;

Todo:

Perfect tracking module; 3/11/2019
Tracking module optimization; 24/11/2019
The matching logic needs to be modified. The current matching logic is just a simple iou calculation;
parameters of optical flow method need to be adjusted;
[x]