Apollo自动驾驶融合了全世界优秀的算法 铃灵狗对车道线感知和高精地图融合的方式产生了好奇 全网没有找到明确的分析和总结 因此亲自阅读源码 找到其中的奥妙之处 模式选择 在 mo
在modules/map/relative_map/conf/relative_map_config.pb.txt文件中对模式进行修改：
lane_source: OFFLINE_GENERATED
存在两种模式：OFFLINE_GENERATED 和 PERCPTION
当使用前者时 高精地图和感知进行融合
使用后者时 只是用感知生成相对地图
本文只讨论前者
文件路径：modules/map/relative_map/navigation_lane.cc
函数名：bool NavigationLane::GeneratePath() 这个是重点分析的函数
下面进入这个函数
在这里进行了模式检查：
if (config_.lane_source() == NavigationLaneConfig::OFFLINE_GENERATED && navigation_line_num > 0)
现在进入这个if中
for (int i = 0; i < navigation_line_num; ++i) {
auto current_navi_path = std::make_shared<NavigationPath>();
auto *path = current_navi_path->mutable_path();
if (ConvertNavigationLineToPath(i, path)) {
current_navi_path->set_path_priority(
navigation_info_.navigation_path(i).path_priority());
navigation_path_list_.emplace_back(
i, default_left_width_, default_right_width_, current_navi_path);
}
}
这里根据导航数据对 navigation_path_list_ 进行了填充 接下来马上进行判断：
if (navigation_path_list_.empty()) {
generate_path_on_perception_func();
return true;
}
如果没有数据 只能使用感知了
如果数据不为空 我们继续
这里对导航路线进行排序 还有设置当前车辆所在路径
其中 current_navi_path_tuple_ 这个变量记录了当前车所在的道路的索引 下面我们将这个变量称为当前车道
//根据车的方向 将导航路线从左到右排序
navigation_path_list_.sort(
[](const NaviPathTuple &left, const NaviPathTuple &right) {
double left_y = std::get<3>(left)->path().path_point(0).y();
double right_y = std::get<3>(right)->path().path_point(0).y();
return left_y > right_y;
});
// Get which navigation path the vehicle is currently on.
double min_d = std::numeric_limits<double>::max();
for (const auto &navi_path_tuple : navigation_path_list_) {
int current_line_index = std::get<0>(navi_path_tuple);
ADEBUG << "Current navigation path index is: " << current_line_index;
double current_d = std::numeric_limits<double>::max();
auto item_iter = last_project_index_map_.find(current_line_index);
if (item_iter != last_project_index_map_.end()) {
current_d = item_iter->second.second;
}
if (current_d < min_d) {
min_d = current_d;
//设置当前车辆所在路径
current_navi_path_tuple_ = navi_path_tuple;
}
}
接下来 将当前车道进行融合 而不是将所有车道进行融合！
//融合感知和导航车道线
// Merge current navigation path where the vehicle is located with perceived
// lane markers.
auto *path = std::get<3>(current_navi_path_tuple_)->mutable_path();
MergeNavigationLineAndLaneMarker(std::get<0>(current_navi_path_tuple_),
path);
进入融合函数
void NavigationLane::MergeNavigationLineAndLaneMarker(const int line_index, common::Path *const path)
进入之后直接进行判断
如果点的数量少于2 则首先生成导航的车道线
if (path->path_point_size() < 2) {
path->Clear();
ConvertNavigationLineToPath(line_index, path);
}
如果还是点的数量过少 那只能使用感知的了 并直接return 不用融合了
//如果路径点还是小于2 则根据感知生成导航车道线
// If the size of "path" points is still smaller than 2, just generate a
// navigation path based on perceived lane markers.
if (path->path_point_size() < 2) {
path->Clear();
ConvertLaneMarkerToPath(perception_obstacles_.lane_marker(), path);
return;
}
common::Path lane_marker_path;
ConvertLaneMarkerToPath(perception_obstacles_.lane_marker(),
&lane_marker_path);
// If the size of lane marker path points is smaller than 2, merging is not
// required.
if (lane_marker_path.path_point_size() < 2) {
return;
}
下面进行融合
前提是只对当前车道进行车道线的融合 因为传入的就只是当前车道
融合原理大概是预先设置一个权重 目前感觉这个权重应该是不变的 这里存个疑
然后遍历导航中这条线每个点 在导航中找到这个点对应的点 根据权重进行融合
代码展示
int lane_marker_index = 0;
double navigation_line_weight = 1.0 - config_.lane_marker_weight();
for (int i = 0; i < path->path_point_size(); ++i) {
auto *point = path->mutable_path_point(i);
double s = point->s();
//计算s距离之后的匹配点 并返回匹配点的索引
//lane_marker_path为感知发来的道路
auto lane_maker_point = GetPathPointByS(lane_marker_path, lane_marker_index,
s, &lane_marker_index);
// For the beginning and ending portions of a navigation path beyond the
// perceived path, only the y-coordinates in the FLU coordinate system are
// used for merging.
const int marker_size = lane_marker_path.path_point_size();
if (lane_marker_index < 0 || lane_marker_index > (marker_size - 1)) {
point->set_y(navigation_line_weight * point->y() +
(1 - navigation_line_weight) * lane_maker_point.y());
lane_marker_index = 0;
continue;
}
*point = common::util::GetWeightedAverageOfTwoPathPoints(
*point, lane_maker_point, navigation_line_weight,
1 - navigation_line_weight);
}
然后回到之前的函数
下面就是各种设置宽度了
先提取了感知的道路宽度
遍历导航中所有的道路（这里不仅仅是处理当前道路）
如果遍历到了当前道路 就设置为感知到的道路宽度
代码展示
//设置导航道路的宽度
// Set the width for the navigation path which the vehicle is currently on.
//注意感知车道线宽度是在这里储存的
double left_width = perceived_left_width_ > 0.0 ? perceived_left_width_
: default_left_width_;
double right_width = perceived_right_width_ > 0.0 ? perceived_right_width_
: default_right_width_;
if (!IsFloatEqual(left_width, default_left_width_) &&
!IsFloatEqual(right_width, default_right_width_)) {
left_width = left_width > default_left_width_ ? left_width - min_d
: left_width + min_d;
right_width = right_width > default_right_width_ ? right_width - min_d
: right_width + min_d;
}
ADEBUG << "The left width of current lane is: " << left_width
<< " and the right width of current lane is: " << right_width;
std::get<1>(current_navi_path_tuple_) = left_width;
std::get<2>(current_navi_path_tuple_) = right_width;
auto curr_navi_path_iter = std::find_if(
std::begin(navigation_path_list_), std::end(navigation_path_list_),
[this](const NaviPathTuple &item) {
return std::get<0>(item) == std::get<0>(current_navi_path_tuple_);
});
if (curr_navi_path_iter != std::end(navigation_path_list_)) {
std::get<1>(*curr_navi_path_iter) = left_width;
std::get<2>(*curr_navi_path_iter) = right_width;
}
// Set the width between each navigation path and its adjacent path.
// The reason for using average of multiple points is to prevent too much
// interference from a singularity.
// If current navigation path is the path which the vehicle is currently
// on, the current lane width uses the perceived width.
//设置每个导航路径与其相邻路径之间的宽度。
//使用多个点的平均值是为了防止奇点产生过多干扰。
//如果当前导航路径是车辆当前所在的路径，则当前车道宽度使用感知宽度。
int average_point_size = 5;
for (auto iter = navigation_path_list_.begin();
iter != navigation_path_list_.end(); ++iter) {
const auto &curr_path = std::get<3>(*iter)->path();
// Left neighbor
auto prev_iter = std::prev(iter);
if (prev_iter != navigation_path_list_.end()) {
const auto &prev_path = std::get<3>(*prev_iter)->path();
average_point_size = std::min(
average_point_size,
std::min(curr_path.path_point_size(), prev_path.path_point_size()));
double lateral_distance_sum = 0.0;
for (int i = 0; i < average_point_size; ++i) {
lateral_distance_sum +=
fabs(curr_path.path_point(i).y() - prev_path.path_point(i).y());
}
double width = lateral_distance_sum /
static_cast<double>(average_point_size) / 2.0;
width = common::math::Clamp(width, config_.min_lane_half_width(),
config_.max_lane_half_width());
auto &curr_left_width = std::get<1>(*iter);
auto &prev_right_width = std::get<2>(*prev_iter);
if (std::get<0>(*iter) == std::get<0>(current_navi_path_tuple_)) {
prev_right_width = 2.0 * width - curr_left_width;
} else {
curr_left_width = width;
prev_right_width = width;
}
}
// Right neighbor
auto next_iter = std::next(iter);
if (next_iter != navigation_path_list_.end()) {
const auto &next_path = std::get<3>(*next_iter)->path();
average_point_size = std::min(
average_point_size,
std::min(curr_path.path_point_size(), next_path.path_point_size()));
double lateral_distance_sum = 0.0;
for (int i = 0; i < average_point_size; ++i) {
lateral_distance_sum +=
fabs(curr_path.path_point(i).y() - next_path.path_point(i).y());
}
double width = lateral_distance_sum /
static_cast<double>(average_point_size) / 2.0;
width = common::math::Clamp(width, config_.min_lane_half_width(),
config_.max_lane_half_width());
auto &curr_right_width = std::get<2>(*iter);
auto &next_left_width = std::get<1>(*next_iter);
if (std::get<0>(*iter) == std::get<0>(current_navi_path_tuple_)) {
next_left_width = 2.0 * width - curr_right_width;
} else {
next_left_width = width;
curr_right_width = width;
}
}
如果权重是动态变化的还请大家评论纠正我
感觉apollo的车道线也就是高精地图和感知通过权重进行融合 没有很特别很精妙的算法 如有不对之处还请指教