C++应用中调用YOLOv3（darknet）进行目标检测

YOLOv3论文：https://pjreddie.com/media/files/papers/YOLOv3.pdf

官网和代码：https://pjreddie.com/darknet/

yolo属于one-stage（检测一步到位），兼顾准确率和速度，特别是最近的v3版本提高了小目标的检测率，是移动端目标检测的热门算法。关于YOLO原理的介绍网上有很多资料请自行百度，本文主要介绍如何在自己的cpp中调用yolov3进行目标检测。

yolo采用自定义的image格式进行图像读取和处理，而一般我们工程中使用较多的是OpenCV或者指向图像数据的指针，因此此处先对图像转换和缩放操作进行修改，代码如下：

#ifndef IMPROCESS_H
#define IMPROCESS_H

#include<opencv2/opencv.hpp>

void imgConvert(const cv::Mat& img, float* dst);

void imgResize(float* src, float* dst,int srcWidth,int srcHeight,int dstWidth,int dstHeight);

void resizeInner(float *src, float* dst,int srcWidth,int srcHeight,int dstWidth,int dstHeight);

#endif // IMPROCESS_H

#include<improcess.h>

void imgConvert(const cv::Mat& img, float* dst){
    uchar *data = img.data;
    int h = img.rows;
    int w = img.cols;
    int c = img.channels();

    for(int k= 0; k < c; ++k){
        for(int i = 0; i < h; ++i){
            for(int j = 0; j < w; ++j){
                dst[k*w*h+i*w+j] = data[(i*w + j)*c + k]/255.;
            }
        }
    }
}

void imgResize(float *src, float* dst,int srcWidth,int srcHeight,int dstWidth,int dstHeight){
    int new_w = srcWidth;
    int new_h = srcHeight;
    if (((float)dstWidth/srcWidth) < ((float)dstHeight/srcHeight)) {
        new_w = dstWidth;
        new_h = (srcHeight * dstWidth)/srcWidth;
    } else {
        new_h = dstHeight;
        new_w = (srcWidth * dstHeight)/srcHeight;
    }

    float* ImgReInner;
    size_t sizeInner=new_w*new_h*3*sizeof(float);
    ImgReInner=(float*)malloc(sizeInner);
    resizeInner(src,ImgReInner,srcWidth,srcHeight,new_w,new_h);

    for(int i=0;i<dstWidth*dstHeight*3;i++){
        dst[i]=0.5;
    }

    for(int k = 0; k < 3; ++k){
        for(int y = 0; y < new_h; ++y){
            for(int x = 0; x < new_w; ++x){
                float val = ImgReInner[k*new_w*new_h+y*new_w+x];
                dst[k*dstHeight*dstWidth + ((dstHeight-new_h)/2+y)*dstWidth + (dstWidth-new_w)/2+x]=val;
            }
        }
    }
    free(ImgReInner);
}

void resizeInner(float *src, float* dst,int srcWidth,int srcHeight,int dstWidth,int dstHeight){
    float* part;
    size_t sizePa=dstWidth*srcHeight*3*sizeof(float);
    part=(float*)malloc(sizePa);

    float w_scale = (float)(srcWidth - 1) / (dstWidth - 1);
    float h_scale = (float)(srcHeight - 1) / (dstHeight - 1);

    for(int k = 0; k < 3; ++k){
        for(int r = 0; r < srcHeight; ++r){
            for(int c = 0; c < dstWidth; ++c){
                float val = 0;
                if(c == dstWidth-1 || srcWidth == 1){
                    val=src[k*srcWidth*srcHeight+r*srcWidth+srcWidth-1];
                } else {
                    float sx = c*w_scale;
                    int ix = (int) sx;
                    float dx = sx - ix;
                    val=(1 - dx) * src[k*srcWidth*srcHeight+r*srcWidth+ix] + dx * src[k*srcWidth*srcHeight+r*srcWidth+ix+1];
                }
                part[k*srcHeight*dstWidth + r*dstWidth + c]=val;
            }
        }
    }

    for(int k = 0; k < 3; ++k){
        for(int r = 0; r < dstHeight; ++r){
            float sy = r*h_scale;
            int iy = (int) sy;
            float dy = sy - iy;
            for(int c = 0; c < dstWidth; ++c){
                float val = (1-dy) * part[k*dstWidth*srcHeight+iy*dstWidth+c];
                dst[k*dstWidth*dstHeight + r*dstWidth + c]=val;
            }
            if(r == dstHeight-1 || srcHeight == 1)
                continue;
            for(int c = 0; c < dstWidth; ++c){
                float val = dy * part[k*dstWidth*srcHeight+(iy+1)*dstWidth+c];
                dst[k*dstWidth*dstHeight + r*dstWidth + c]+=val;
            }
        }
    }
    free(part);
}

其中，imgConvert函数将OpenCV的图像由RGBRGBRGB...转化为yolo的RRRGGGBBB...格式（由代码可知，yolo输入图像的像素取值范围为0~1）。imgResize函数将图像缩放到cfg指定的网络输入的大小。代码修改自yolo的源码，将其image格式改为我们需要的指针形式。

接下来是调用darknet的代码，为了让代码跑通，我们首先用OpenCV读取视频，然后将OpenCV的图像转为指针指向的数据格式（如果想直接采用OpenCV可自行修改）。代码如下：

#include<iostream>
#include<opencv2/opencv.hpp>
#include<darknet.h>
#include<improcess.h>

using namespace std;
using namespace cv;

float colors[6][3] = { {1,0,1}, {0,0,1},{0,1,1},{0,1,0},{1,1,0},{1,0,0} };

float get_color(int c, int x, int max){
    float ratio = ((float)x/max)*5;
    int i = floor(ratio);
    int j = ceil(ratio);
    ratio -= i;
    float r = (1-ratio) * colors[i][c] + ratio*colors[j][c];
    return r;
}


int main()
{
    string cfgfile = "/home/chnn/darknet/cfg/yolov3.cfg";//读取模型文件，请自行修改相应路径
    string weightfile = "/home/chnn/darknet/yolov3.weights";
    float thresh=0.5;//参数设置
    float nms=0.35;
    int classes=80;

    network *net=load_network((char*)cfgfile.c_str(),(char*)weightfile.c_str(),0);//加载网络模型
    set_batch_network(net, 1);
    VideoCapture capture("/home/chnn/video/videoCapture6.mp4");//读取视频，请自行修改相应路径
    Mat frame;
    Mat rgbImg;

    vector<string> classNamesVec;
    ifstream classNamesFile("/home/chnn/darknet/data/coco.names");//标签文件coco有80类

    if (classNamesFile.is_open()){
        string className = "";
        while (getline(classNamesFile, className))
            classNamesVec.push_back(className);
    }

    bool stop=false;
    while(!stop){
        if (!capture.read(frame)){
            printf("fail to read.\n");
            return 0;
        }
        cvtColor(frame, rgbImg, cv::COLOR_BGR2RGB);

        float* srcImg;
        size_t srcSize=rgbImg.rows*rgbImg.cols*3*sizeof(float);
        srcImg=(float*)malloc(srcSize);

        imgConvert(rgbImg,srcImg);//将图像转为yolo形式

        float* resizeImg;
        size_t resizeSize=net->w*net->h*3*sizeof(float);
        resizeImg=(float*)malloc(resizeSize);
        imgResize(srcImg,resizeImg,frame.cols,frame.rows,net->w,net->h);//缩放图像

        network_predict(net,resizeImg);//网络推理
        int nboxes=0;
        detection *dets=get_network_boxes(net,rgbImg.cols,rgbImg.rows,thresh,0.5,0,1,&nboxes);

        if(nms){
            do_nms_sort(dets,nboxes,classes,nms);
        }

        vector<cv::Rect>boxes;
        boxes.clear();
        vector<int>classNames;

        for (int i = 0; i < nboxes; i++){
            bool flag=0;
            int className;
            for(int j=0;j<classes;j++){
                if(dets[i].prob[j]>thresh){
                    if(!flag){
                        flag=1;
                        className=j;
                    }
                }
            }
            if(flag){
                int left = (dets[i].bbox.x - dets[i].bbox.w / 2.)*frame.cols;
                int right = (dets[i].bbox.x + dets[i].bbox.w / 2.)*frame.cols;
                int top = (dets[i].bbox.y - dets[i].bbox.h / 2.)*frame.rows;
                int bot = (dets[i].bbox.y + dets[i].bbox.h / 2.)*frame.rows;

                if (left < 0)
                    left = 0;
                if (right > frame.cols - 1)
                    right = frame.cols - 1;
                if (top < 0)
                    top = 0;
                if (bot > frame.rows - 1)
                    bot = frame.rows - 1;

                Rect box(left, top, fabs(left - right), fabs(top - bot));
                boxes.push_back(box);
                classNames.push_back(className);
            }
        }
        free_detections(dets, nboxes);

        for(int i=0;i<boxes.size();i++){
            int offset = classNames[i]*123457 % 80;
            float red = 255*get_color(2,offset,80);
            float green = 255*get_color(1,offset,80);
            float blue = 255*get_color(0,offset,80);

            rectangle(frame,boxes[i],Scalar(blue,green,red),2);

            String label = String(classNamesVec[classNames[i]]);
            int baseLine = 0;
            Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
            putText(frame, label, Point(boxes[i].x, boxes[i].y + labelSize.height),
                    FONT_HERSHEY_SIMPLEX, 1, Scalar(red, blue, green),2);
        }
        imshow("video",frame);

        int c=waitKey(30);
              if((char)c==27)
                  break;
              else if(c>=0)
                  waitKey(0);

        free(srcImg);
        free(resizeImg);
    }
    free_network(net);
    capture.release();
    return 1;
}

链接上darknet的动态库并让代码运行，最后得出来的结果应该是这样的：

请忽略人体姿态信息。。。