Pre Merge pull request !23 from Daniel/lxm

algorithm/veri/cuda/veri_det_cuda_impl.cpp

@@ -74,10 +74,10 @@ namespace sv
   bool VeriDetectorCUDAImpl::cudaSetup()
   {
 #ifdef WITH_CUDA
-    std::string trt_model_fn = get_home() + SV_MODEL_DIR + "model.engine";
+    std::string trt_model_fn = get_home() + SV_MODEL_DIR + "veri.engine";
     if (!is_file_exist(trt_model_fn))
     {
-      throw std::runtime_error("SpireCV (104) Error loading the VERI TensorRT model (File Not Exist)");
+      throw std::runtime_error("SpireCV (104) Error loading the LandingMarker TensorRT model (File Not Exist)");
     }
     char *trt_model_stream{nullptr};
     size_t trt_model_size{0};
@@ -107,7 +107,7 @@ namespace sv
 
     delete[] trt_model_stream;
     const ICudaEngine &cu_engine = this->_trt_context->getEngine();
-    assert(cu_engine.getNbBindings() == 2);
+    assert(cu_engine.getNbBindings() == 3);
 
     this->_input_index = cu_engine.getBindingIndex("input");
     this->_output_index1 = cu_engine.getBindingIndex("output");
@@ -123,7 +123,6 @@ namespace sv
     this->_p_data = new float[2 * 3 * 224 * 224];
     this->_p_prob1 = new float[2 * 576];
     this->_p_prob2 = new float[2 * 1280];
-    this->_p_prob3 = new float[2 * 1280];
     // Input
     TRTCHECK(cudaMemcpyAsync(_p_buffers[this->_input_index], this->_p_data, 2 * 3 * 224 * 224 * sizeof(float), cudaMemcpyHostToDevice, this->_cu_stream));
     // this->_trt_context->enqueue(1, _p_buffers, this->_cu_stream, nullptr);
@@ -139,11 +138,12 @@ namespace sv
 
   void VeriDetectorCUDAImpl::cudaRoiCNN(
       std::vector<cv::Mat> &input_rois_,
-      std::vector<int> &output_labels_)
+      std::vector<float> &output_labels_)
   {
 #ifdef WITH_CUDA
 
-    for (int i = 0; i < 2; ++i)
+
+    for (int i = 0; i < 2; i++)
     {
       for (int row = 0; row < 224; ++row)
       {
@@ -151,14 +151,15 @@ namespace sv
         for (int col = 0; col < 224; ++col)
         {
           // mean=[136.20, 141.50, 145.41], std=[44.77, 44.20, 44.30]
-          this->_p_data[224 * 224 * 3 * i + col + row * 224] = ((float)uc_pixel[0] - 136.20f) / 44.77f;
+          this->_p_data[col + row * 224 + 224 * 224 * 3 * i] = ((float)uc_pixel[0] - 136.20f) / 44.77f;
-          this->_p_data[224 * 224 * 3 * i + col + row * 224 + 224 * 224] = ((float)uc_pixel[1] - 141.50f) / 44.20f;
+          this->_p_data[col + row * 224 + 224 * 224  + 224 * 224 * 3 * i] = ((float)uc_pixel[1] - 141.50f) / 44.20f;
-          this->_p_data[224 * 224 * 3 * i + col + row * 224 + 224 * 224 * 2] = ((float)uc_pixel[2] - 145.41f) / 44.30f;
+          this->_p_data[col + row * 224 + 224 * 224 * 2  + 224 * 224 * 3 * i] = ((float)uc_pixel[2] - 145.41f) / 44.30f;
           uc_pixel += 3;
         }
       }
     }
 
+
     // Input
     TRTCHECK(cudaMemcpyAsync(_p_buffers[this->_input_index], this->_p_data, 2 * 3 * 224 * 224 * sizeof(float), cudaMemcpyHostToDevice, this->_cu_stream));
     // this->_trt_context->enqueue(1, _p_buffers, this->_cu_stream, nullptr);
@@ -180,10 +181,9 @@ namespace sv
       }
     }
 
-    // 计算两个数组的余弦相似性
     float similarity = cosineSimilarity(_p_prob2, _p_prob2 + 1280, 1280);
-    std::cout << "余弦相似性: " << similarity << std::endl;
+    output_labels_.push_back(label);
-    std::cout << "VERI LABEL: " << label << std::endl;
+    output_labels_.push_back(similarity);
   }
 #endif
 }

algorithm/veri/cuda/veri_det_cuda_impl.h

@@ -29,14 +29,13 @@ public:
   bool cudaSetup();
   void cudaRoiCNN(
     std::vector<cv::Mat>& input_rois_,
-    std::vector<int>& output_labels_
+    std::vector<float>& output_labels_
   );
 
 #ifdef WITH_CUDA
   float *_p_data;
   float *_p_prob1;
   float *_p_prob2;
-  float *_p_prob3;
   nvinfer1::IExecutionContext *_trt_context;
   int _input_index;
   int _output_index1;

algorithm/veri/sv_veri_det.cpp

@@ -7,55 +7,67 @@
 #include "veri_det_cuda_impl.h"
 #endif
 
-
+namespace sv
-namespace sv {
-
-
-VeriDetector::VeriDetector()
 {
-  this->_cuda_impl = new VeriDetectorCUDAImpl;
-}
-VeriDetector::~VeriDetector()
-{
-}
 
-bool VeriDetector::setupImpl()
+  VeriDetector::VeriDetector()
-{
+  {
-#ifdef WITH_CUDA
+    this->_cuda_impl = new VeriDetectorCUDAImpl;
-  return this->_cuda_impl->cudaSetup();
+  }
-#endif
+  VeriDetector::~VeriDetector()
-  return false;
-}
-
-void VeriDetector::roiCNN(
-  std::vector<cv::Mat>& input_rois_,
-  std::vector<int>& output_labels_
-)
-{
-#ifdef WITH_CUDA
-  this->_cuda_impl->cudaRoiCNN(
-    input_rois_,
-    output_labels_
-  );
-#endif
-}
-
-
-
-void VeriDetector::detect(cv::Mat img1_, cv::Mat img2_, TargetsInFrame& tgts_)
-{
-  if (!_params_loaded)
   {
-    this->_load();
-    this->_loadLabels();
-    _params_loaded = true;
   }
 
-  std::vector<cv::Mat> e_roi = {img1_, img2_};
+  bool VeriDetector::setupImpl()
+  {
+#ifdef WITH_CUDA
+    return this->_cuda_impl->cudaSetup();
+#endif
+    return false;
+  }
 
-  std::vector<int> output_labels;
+  void VeriDetector::roiCNN(
-  roiCNN(e_roi, output_labels);
+      std::vector<cv::Mat> &input_rois_,
-}
+      std::vector<float> &output_labels_)
+  {
+#ifdef WITH_CUDA
+    this->_cuda_impl->cudaRoiCNN(
+        input_rois_,
+        output_labels_);
+#endif
+  }
+
+  void VeriDetector::detect(cv::Mat img1_, cv::Mat img2_, TargetsInFrame &tgts_)
+  {
+    if (!_params_loaded)
+    {
+      this->_load();
+      this->_loadLabels();
+      _params_loaded = true;
+    }
+
+    std::vector<cv::Mat> input_rois_ = {img1_, img2_};
+
+#ifdef WITH_CUDA
+    std::vector<float> output_labels;
+    roiCNN(input_rois_, output_labels);
+#endif
+
+    tgts_.setSize(img1_.cols, img1_.rows);
+    tgts_.setFOV(this->fov_x, this->fov_y);
+    auto t1 = std::chrono::system_clock::now();
+    tgts_.setFPS(1000.0 / std::chrono::duration_cast<std::chrono::milliseconds>(t1 - this->_t0).count());
+    this->_t0 = std::chrono::system_clock::now();
+    tgts_.setTimeNow();
+
+     if (output_labels.size() > 0)
+    {
+        Target tgt;
+
+        tgt.category_id = output_labels[0];
+        tgt.score = output_labels[1];
+        tgts_.targets.push_back(tgt);
+    }
+  }
 
 }
-

include/sv_veri_det.h

@@ -25,7 +25,7 @@ protected:
   bool setupImpl();
   void roiCNN(
     std::vector<cv::Mat>& input_rois_,
-    std::vector<int>& output_labels_
+    std::vector<float>& output_labels_
   );
 
   VeriDetectorCUDAImpl* _cuda_impl;

samples/demo/veri.cpp

@@ -5,40 +5,66 @@
 
 using namespace std;
 
-int main(int argc, char *argv[]) {  
+int main(int argc, char *argv[])
-  // 打开摄像头
+{
+  // 实例化 圆形降落标志 检测器类
   sv::VeriDetector veri;
+  // 手动导入相机参数，如果使用Amov的G1等吊舱或相机，则可以忽略该步骤，将自动下载相机参数文件
   veri.loadCameraParams(sv::get_home() + "/SpireCV/calib_webcam_640x480.yaml");
 
-  cv::VideoCapture cap1("/home/amov/Videos/com/FlyVideo_2023-09-02_11-36-00.avi");
+  // 打开摄像头
-  cv::VideoCapture cap2("/home/amov/Videos/com/FlyVideo_2023-09-02_11-41-55.avi");
+  sv::Camera cap1,cap2;
-  // cap.setWH(640, 480);
+  cap1.setWH(640, 480);
-  // cap.setFps(30);
+  cap1.setFps(30);
-  //cap.open(sv::CameraType::WEBCAM, 0);  // CameraID 0
+  cap1.open(sv::CameraType::WEBCAM, 4);  // CameraID 0
+  
+  cap2.setWH(640, 480);
+  cap2.setFps(30);
+  cap2.open(sv::CameraType::WEBCAM, 10);  // CameraID 0
+
+
+
   // 实例化OpenCV的Mat类，用于内存单帧图像
-
+  cv::Mat img1, img2;
-
-  cv::Mat img1,img2;
   int frame_id = 0;
+  int w = 224;
+
   while (1)
   {
     // 实例化SpireCV的 单帧检测结果 接口类 TargetsInFrame
     sv::TargetsInFrame tgts(frame_id++);
-
     // 读取一帧图像到img
     cap1.read(img1);
     cap2.read(img2);
 
-    cv::resize(img1, img1, cv::Size(224, 224));
+    cv::resize(img1, img1, cv::Size(w, w));
-    cv::resize(img2, img2, cv::Size(224, 224));
+    cv::resize(img2, img2, cv::Size(w, w));
 
+    // 执行 降落标志 检测
     veri.detect(img1, img2, tgts);
+    // 可视化检测结果，叠加到img上
+    // sv::drawTargetsInFrame(img, tgts);
 
-    
+    // // 控制台打印 降落标志 检测结果
+    printf("Frame-[%d]\n", frame_id);
+    // 打印当前检测的FPS
+    printf("  FPS = %.2f\n", tgts.fps);
+    // 打印当前相机的视场角（degree）
+    printf("  FOV (fx, fy) = (%.2f, %.2f)\n", tgts.fov_x, tgts.fov_y);
+    // 打印当前输入图像的像素宽度和高度
+    printf("  Frame Size (width, height) = (%d, %d)\n", tgts.width, tgts.height);
+    for (int i = 0; i < tgts.targets.size(); i++)
+    {
 
-    // 显示img
+      // 打印每个 降落标志 的置信度
-    // cv::imshow("img", img);
+      printf("  Similarity Score = %.3f\n", tgts.targets[i].score);
-    // cv::waitKey(10);
+      printf("  Category ID = %d\n", tgts.targets[i].category_id);
+    }
+
+    // 显示检测结果img
+    cv::imshow("img1", img1);
+    cv::imshow("img2", img2);
+    cv::waitKey(10);
   }
 
   return 0;