object_filter_bank

目录

object_filter_bank.cc

Filter 

roi_boundary_filter

Filter

BuildWorldPolygons

FillObjectRoiFlag

FilterObjectsOutsideBoundary

FilterObjectsInsideBoundary

FilterObjectsByConfidence

---

 

清空存储filter的成员变量，然后遍历config文件中的filter_name, 从这里看配置文件里面就一个 ROIBoundaryFilter（有需要可以自己添加），然后通过registerer实例化

PERCEPTION_REGISTER_OBJECTFILTER(ROIBoundaryFilter);

初始化filter，然后push_back到filter_bank_, 当然因为配置文件里面就一个，所以遍历结束。

filter_bank_.clear();
  for (int i = 0; i < config.filter_name_size(); ++i) {
    const auto& name = config.filter_name(i);
    BaseObjectFilter* filter =
        BaseObjectFilterRegisterer::GetInstanceByName(name);
    if (!filter) {
      AINFO << "Failed to find object filter: " << name << ", skipped";
      continue;
    }
    if (!filter->Init()) {
      AINFO << "Failed to init object filter: " << name << ", skipped";
      continue;
    }
    filter_bank_.push_back(filter);
    AINFO << "Filter bank add filter: " << name;
  }

object_filter_bank.cc

Filter 

依次执行各个filter。

for (auto& filter : filter_bank_) {
    if (!filter->Filter(options, frame)) {
      AINFO << "Failed to filter objects in: " << filter->Name();
    }
  }

按照代码的结构，应该是方便添加更多的filter。目前就只有一个ROIBoundaryFilter，所以开始执行ROIBoundaryFilter。

roi_boundary_filter

配置文件

 成员变量

// params
  double distance_to_boundary_threshold_ = 1.0;
  double inside_threshold_ = 1.0;
  float confidence_threshold_ = 0.5f;
  float cross_roi_threshold_ = 0.6f;

Filter

先是检查frame中的数据，包含object，hdmap_struct. 区分障碍物前后景来判断object是否在ROI。

然后依次执行

FillObjectRoiFlag(options, frame);
BuildWorldPolygons(options, *frame);
if (distance_to_boundary_threshold_ >= 0.f) {
    FilterObjectsOutsideBoundary(options, frame, &objects_valid_flag_);
  }
  if (inside_threshold_ >= 0.f) {
    FilterObjectsInsideBoundary(options, frame, &objects_valid_flag_);
  }
FilterObjectsByConfidence(options, frame, &objects_valid_flag_);

最后更新一下objects size。

BuildWorldPolygons

目前ObjectFilterOptions为空 

const Eigen::Affine3d& pose = frame.lidar2world_pose;

for (size_t j = 0; j < obj->polygon.size(); ++j) {
        local_point(0) = obj->polygon[j].x;
        local_point(1) = obj->polygon[j].y;
        local_point(2) = obj->polygon[j].z;
        world_point = pose * local_point;
        polygons_in_world_[i][j].x = world_point(0);
        polygons_in_world_[i][j].y = world_point(1);
        polygons_in_world_[i][j].z = world_point(2);

FillObjectRoiFlag

这里回答了一个问题，什么是cross？就是object有部分点云在ROI中，这就算cross。

如果object有一个点云在ROI中，就算这个object.is_in_roi。

cross就是有部分点云在ROI中，这里有个问题，假如num_points_in_roi = cloud.size ?

计算一个in roi点云占比 ratio。

如果 ratio小于阈值，或者  有部分点云在ROI中并且confidence<= .11f. 这样object就算 cross_roi。

总结，这里把object分成了is_in_roi和objects_cross_roi两种，并且cross_roi也算is_in_roi。

obj->lidar_supplement.is_in_roi =
        obj->lidar_supplement.num_points_in_roi > 0;
    bool cross = (obj->lidar_supplement.num_points_in_roi !=
                  obj->lidar_supplement.cloud.size());
    float ratio = static_cast<float>(obj->lidar_supplement.num_points_in_roi) /
                  static_cast<float>(obj->lidar_supplement.cloud.size());
    if (ratio < cross_roi_threshold_  // a hacked minimum support value
        || (cross && obj->confidence <= .11f)) {
      objects_cross_roi_[i] = true;

FilterObjectsOutsideBoundary

前提条件，得确保object在roi外边

if (!obj->lidar_supplement.is_in_roi)

遍历obstacle的每一个点云

for (auto& point : polygons_in_world_[i].points()) {
        dist_to_boundary = 0.0;
        min_dist_to_boundary = kDoubleMax;
        world_point << point.x, point.y, point.z;

 然后计算点到每一个boundary的距离

for (const auto& boundary : road_boundary) {
          perception::common::CalculateDistAndDirToBoundary(
              world_point, boundary.left_boundary, boundary.right_boundary,
              &dist_to_boundary, &direction);

用min_dist_to_boundary来保存最小的距离 

if (min_dist_to_boundary > dist_to_boundary) {
            min_dist_to_boundary = dist_to_boundary;
          }

如果最小距离是小于阈值的，则置flag为true。 

if (min_dist_to_boundary <= distance_to_boundary_threshold_) {
    (*objects_valid_flag)[i] = true;
    break;
}

FilterObjectsInsideBoundary

前提条件，确保object in roi 

if (obj->lidar_supplement.is_in_roi && !objects_cross_roi_[i] &&
        obj->confidence <= .11f) 

其余步骤同上，最后的判断条件相反，因为这是是在里面，所以越里面越好。

if (min_dist_to_boundary > inside_threshold_) {
     (*objects_valid_flag)[i] = true;
     break;
}

FilterObjectsByConfidence

将那些cross roi但是confidence低于阈值的障碍物 object_valid_flag置为false，应该就是过滤掉了。

for (size_t i = 0; i < objects.size(); ++i) {
    if (objects_cross_roi_[i] || !objects[i]->lidar_supplement.is_in_roi) {
      if (objects[i]->confidence < confidence_threshold_) 
            (*objects_valid_flag)[i] = false;
    }
}