实时视频人脸识别尝试（使用ShiqiYu老师的人脸识别库）

使用ShiqiYu老师的 face landmark detection方法进行人脸识别与人脸转动角度的计算，使用的是 ShiqiYu老师在2017年1月更新的代码。

Github地址：https://github.com/ShiqiYu/libfacedetection

其代码较2015年的版本做了一些修改， 其中提供了 frontal,frontal-surveillance,multiview,multiview_reinforce四种方法。每种方法的耗费时间与检测的人脸转向的范围有所不同，在Github中对其效果差异做了详细说明。

使用multiview_reinforce为例进行编写，此次改动在调用函数中做了参数的修改需要注意。

各个参数的含义如下：

FACEDETECTDLL_API int * facedetect_multiview_reinforce(unsigned char * result_buffer, //buffer memory for storing face detection results, !!its size must be 0xC004 Bytes!!

unsigned char * gray_image_data, int width, int height, int step, //input image, it must be gray (single-channel) image!

float scale, //scale factor for scan windows

int min_neighbors, //how many neighbors each candidate rectangle should have to retain it

int min_object_width, //Minimum possible face size. Faces smaller than that are ignored.

int max_object_width = 0, //Maximum possible face size. Faces larger than that are ignored. It is the largest posible when max_object_width=0.

int doLandmark = 0); // landmark detection

在提供的example例子上进行了修改，使用OpenCV接口读取摄像头的图片，进行实时操作。代码如下：

//2017.2.6 
  
 #include <stdio.h> 
 #include <opencv2/opencv.hpp> 
 #include "facedetect-dll.h" 
  
 #pragma comment(lib,"libfacedetect.lib") 
 //#pragma comment(lib,"libfacedetect-x64.lib") 
  
 //define the buffer size. Do not change the size! 
 #define DETECT_BUFFER_SIZE 0x20000 
 using namespace cv; 
  
 void OnMouseAction(int event, int x, int y, int falgs, void *ustc); 
 bool stime = false; 
  
 int main(int argc, char* argv[]) 
 { 
  
 	Mat gray; 
 	Mat frame; 
 	Mat new_frame; 
 	int resize_height = 480; 
 	int resize_width = 640; 
  
 	int * pResults = NULL;  
  
     unsigned char * pBuffer = (unsigned char *)malloc(DETECT_BUFFER_SIZE); 
     if(!pBuffer) 
     { 
         fprintf(stderr, "Can not alloc buffer.\n"); 
         return -1; 
     } 
 	VideoCapture capture(0); 
  
 	while (stime==false) 
 	{ 
 		capture >> frame; 
 		cv::resize(frame, new_frame, cv::Size(resize_width, resize_height), (0.0), (0.0), cv::INTER_LINEAR); 
 		cvtColor(new_frame, gray, CV_BGR2GRAY); 
 		int doLandmark = 1; 
  
 		pResults = facedetect_multiview_reinforce(pBuffer, (unsigned char*)(gray.ptr(0)), gray.cols, gray.rows, (int)gray.step, 
 			1.2f, 3, 24, 0, doLandmark); 
  
 		//printf("%d faces detected.\n", (pResults ? *pResults : 0)); 
 		Mat result_multiview_reinforce = new_frame.clone();; 
 		//print the detection results 
 		for (int i = 0; i < (pResults ? *pResults : 0); i++) 
 		{ 
 			short * p = ((short*)(pResults + 1)) + 142 * i; 
 			int x = p[0]; 
 			int y = p[1]; 
 			int w = p[2]; 
 			int h = p[3]; 
 			int neighbors = p[4]; 
 			int angle = p[5]; 
  
 			printf("face_rect=[%d, %d, %d, %d], neighbors=%d, angle=%d\n", x, y, w, h, neighbors, angle); 
 			rectangle(result_multiview_reinforce, Rect(x, y, w, h), Scalar(0, 255, 0), 2); 
 			if (doLandmark) 
 			{ 
 				for (int j = 0; j < 68; j++) 
 					circle(result_multiview_reinforce, Point((int)p[6 + 2 * j], (int)p[6 + 2 * j + 1]), 1, Scalar(0, 255, 0)); 
 			} 
 		} 
 		imshow("Results_multiview_reinforce", result_multiview_reinforce); 
 		waitKey(30); 
 		setMouseCallback("Results_multiview_reinforce",OnMouseAction); 
 	} 
     //release the buffer 
     free(pBuffer); 
  
 	return 0; 
 } 
  
 void OnMouseAction(int event, int x, int y, int falgs, void *ustc) 
 { 
 	if (event==CV_EVENT_LBUTTONDOWN) 
 	{ 
 		stime = true; 
 	}  
 	else 
 	{ 
 		stime = false;  
 	} 
 } 

效果图如下：

评价一下性能的话，识别效果确实不错。不足点的话一是当人脸有部分遮挡、低头或仰头时无法识别，因此更适合应用于正面下人脸左右摆动识别，如各种身份确认等领域。不是用在安防等领域的检测跟踪的应用背景中。