【電腦視覺】步態能量圖GEI

步態能量圖（Gait Engery Image, GEI）是步態檢測中最非常常用的特徵，提取方法簡單，也能很好的表現步態的速度，形態等特徵。其定義如下：

其中，表示在第q個步態序列中，時刻t的步態剪影圖中座標為(x,y)的像素值。

步態周期的判斷使用步態剪影的寬、高之比即可，這個值比較容易而且隨步態狀態呈現周期性變化。

單張步態剪影圖需調節成寬為W，高為H的大小。調節時保持剪影的比例不變，即如果剪影本身w'/h'<W/H，則將剪影放縮為W*(W*h'/w')大小，並在W*H豎直置中放置。

得到rescaled的步態剪影的代碼：

// get resized gait imageif(!walk_img.empty()){vector<vector<Point> > contours;vector<Vec4i> hierarchy;Mat walk_img_tmp;threshold(walk_img,walk_img_tmp,128,255,THRESH_BINARY);findContours( walk_img_tmp, contours, hierarchy, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE, Point(0, 0) );vector<vector<Point> > contours_poly( contours.size() );vector<Rect> boundRect( contours.size() );int maxRectHeight=0;int maxRectId=0;if(contours.size()>0){for( int i = 0; i< contours.size(); i++ ){//drawContours( walk_img, contours, i, Scalar(255,255,255), 2, 8, hierarchy, 0, Point() );//Approximates a polygonal curve(s) with the specified precision.approxPolyDP( Mat(contours[i]), contours_poly[i], 3, true );//Calculates the up-right bounding rectangle of a point set.boundRect[i] = boundingRect( Mat(contours_poly[i]) );if(boundRect[i].height>maxRectHeight){maxRectHeight = boundRect[i].height;maxRectId = i;}}//rectangle( walk_img, boundRect[maxRectId].tl(), boundRect[maxRectId].br(), Scalar(255,255,255), 2, 8, 0 );double aspect_ratio=(double)boundRect[maxRectId].height/boundRect[maxRectId].width;double base_aspect_ratio=(double)gei_height/gei_width;aspect_ratios.push_back(aspect_ratio);if(aspect_ratio>=base_aspect_ratio){Mat gait_roi=walk_img(boundRect[maxRectId]);Mat gait_roi_tmp;double resize_scale=double(gei_height)/gait_roi.rows;resize(gait_roi,gait_roi_tmp,Size(),resize_scale,resize_scale);Mat gait_img=Mat::zeros(gei_height,gei_width,CV_8UC1);for(int i=0;i<gei_height;i++){uchar* p_tmp=gait_roi_tmp.ptr<uchar>(i);uchar* p=gait_img.ptr<uchar>(i);for(int j=(gei_width-gait_roi_tmp.cols)/2,k=0;k<gait_roi_tmp.cols;k++,j++){p[j]=p_tmp[k];}}gait_imgs.push_back(gait_img);}else{Mat gait_roi=walk_img(boundRect[maxRectId]);Mat gait_roi_tmp;double resize_scale=double(gei_width)/gait_roi.cols;resize(gait_roi,gait_roi_tmp,Size(),resize_scale,resize_scale);Mat gait_img=Mat::zeros(gei_height,gei_width,CV_8UC1);int i=(gei_height-gait_roi_tmp.rows)/2;for(int k=0;k<gait_roi_tmp.rows;k++,i++){uchar* p_tmp=gait_roi_tmp.ptr<uchar>(k);uchar* p=gait_img.ptr<uchar>(i);for(int j=0;j<gei_width;j++){p[j]=p_tmp[j];}}gait_imgs.push_back(gait_img);}}

得到GEI即把上一步每個周期得到的所有圖加權平均即可。

if(aspect_ratios.size()<4)break;// get gait feature: gait energy imagevector<int> max_ids;for(int i=2;i<aspect_ratios.size()-2;i++){if((aspect_ratios[i]>aspect_ratios[i-1])&&(aspect_ratios[i]>aspect_ratios[i-2])&&(aspect_ratios[i]>aspect_ratios[i+1])&&(aspect_ratios[i]>aspect_ratios[i+2]))max_ids.push_back(i);}// for all gait cyclesfor(int cycle_id=1;cycle_id<max_ids.size();cycle_id++){int gait_start_id = max_ids[cycle_id-1];int gait_end_id = max_ids[cycle_id]-1;Mat gait_energy_img=Mat::zeros(gei_height,gei_width, CV_32F);if(gait_end_id-gait_start_id>=6 && gait_end_id-gait_start_id<30){for(int g=gait_start_id;g<=gait_end_id;g++){Mat gait=gait_imgs[g];Mat gait_tmp;gait.convertTo(gait_tmp,CV_32F);gait_energy_img = gait_energy_img+gait_tmp;#ifdef GAIT_DEBUGchar tmp[50];itoa(g,tmp,10);imshow(tmp,gait);#endif}//waitKey(10000);gait_energy_img = gait_energy_img/(float)(gait_end_id-gait_start_id+1);for(int r=0;r<gait_energy_img.rows;r++){float* p=gait_energy_img.ptr<float>(r);for(int c=0;c<gait_energy_img.cols;c++)feature_out<<p[c]<<" ";}feature_out<<endl;label_out<<people_id_iter<<" "<<walk_condition_iter_iter<<endl;cout<<"gait feature cycle #"<<cycle_id-1<<endl;#ifdef GAIT_DEBUGMat gait_enery_img_show;gait_energy_img.convertTo(gait_enery_img_show,CV_8UC1);imshow("GEI",gait_enery_img_show);waitKey(10000);#endif}}}

在CASIA Dataset B 資料集上得到每個角度GEI圖：