查看原文
其他

技术文档|规划模块中的路径边界决策

阿波君 Apollo开发者社区 2022-07-29


从开发套件出发,手把手教你如何实践Apollo自动驾驶技术,给对自动驾驶研究感兴趣的开发者带来实质帮助,加速每一位开发者的研发进程。本文将介绍 规划模块中的路径边界决策,主要从以下几点为大家进行详细讲解。

  • 概览

  • 路径边界决策代码及对应版

    • 类关系

    • 路径边界决策数据

  • 路径边界决策代码流程及框架

  • 路径边界决策算法解析

    • Fallback

    • Pull over

    • Lane change

    • Regular



路径边界决策是规划模块的任务,属于Task中的Decider类别。

规划模块的运动总体流程图如下:



总体流程图以lane follow场景为例进行说明,这里只说明主体流程,不涉及所有细节。Task的主要功能位于Process函数中。


  • 第一,规划模块的入口函数是PlanningComponent的Proc

  • 第二,以规划模式OnLanePlanning,执行RunOnce在RunOnce中先执行交通规则,再规划轨迹,规划轨迹的函数是Plan

  • 第三,进入到PublicRoadPlanner中的Plan函数,需进行轨迹规划。ScenarioManager的Update函数根据当前的scenario_type选择合适的场景,这里的流程图是以lane follow为例。

  • 第四,选择lane follow的场景后,执Process函数,执行LaneFollowStage中的Process函数,在PlanOnReferenceLine中执行LaneFollowStage中的所有的Task。通过调用Excute函数执行Task,Excute调用了Task的Process(以Decider为例子)函数。在上图的最后一个方框内,TaskType指的不是具体的类名称,而是代表所有的Task类型。虚线的箭头,表示在LaneFollowStage中按照Vector中的顺序执行所有的任务。

  • 第五,Task的流程都在Process函数中,对Task的讲解都从Process函数开始。



本节说明path_bounds_decider任务,

可参考Apollo r6.0.0path_bounds_decider


类关系



(1)继承关系

  • PathBoundsDecider类继承Decider类,实现了Process方法,路径边界决策主要的执行过程在Process方法中。


// modules/planning/tasks/deciders/path_bounds_decider/path_bounds_decider.hclass PathBoundsDecider : public Decider {... };

  • Decider类继承Task类,实现类Excute方法,主要是给两个变量赋值:Framereference_line_info,并执行Process方法,对应上述的Process方法。


// modules/planning/tasks/deciders/decider.hclass Decider : public Task {... };// modules/planning/tasks/deciders/decider.ccapollo::common::Status Decider::Execute( Frame* frame, ReferenceLineInfo* reference_line_info) { Task::Execute(frame, reference_line_info); return Process(frame, reference_line_info);}apollo::common::Status Decider::Execute(Frame* frame) { Task::Execute(frame); return Process(frame);}


Task类:定义类保护类型的变量,是路径边界决策的输入。


// modules/planning/tasks/task.hclass Task { public: // 虚方法,主要是给frame和reference_line_info赋值 virtual common::Status Execute(Frame* frame, ReferenceLineInfo* reference_line_info); virtual common::Status Execute(Frame* frame);
protected: // frame和reference_line_info变量 Frame* frame_ = nullptr; ReferenceLineInfo* reference_line_info_ = nullptr;
// 配置与名字 TaskConfig config_; std::string name_;... };

(2)调用

主要描述Task在Stage中是如何创建和调用的。

TaskFactory类:注册所有的Task,包括Decider、Optimizer和Other(E2E的Task)。


// modules/planning/tasks/task_factory.hclass TaskFactory { public: // 两个函数都是static属性 static void Init(...); // 在初始化函数中,注册所有的task static std::unique_ptr<Task> CreateTask(...); // 创建具体task的实例,返回指向该实例的指针... };

Stage中Task的创建与执行

  • 创建:在Stage的构造函数中根据Stage配置创建Task,并将指针放入到Task_Task_list_中。

  • 使用:在具体的Stage中,重写Process方法,调用Process方法,进而调用ExecuteTask*方法

    (ExecuteTaskOnReferenceLine),最后调用相应的Task的Process方法。

// modules/planning/scenarios/stage.hclass Stage { // 在构造函数中根据stage的配置创建task Stage(const ScenarioConfig::StageConfig& config, const std::shared_ptr<DependencyInjector>& injector); public:
// 纯虚函数,留给具体的stage实现,不同的stage有不同的实现逻辑 virtual StageStatus Process( const common::TrajectoryPoint& planning_init_point, Frame* frame) = 0; protected: // 三个执行task的函数,在每个函数中都调用类task的Excute方法,进一步调用具体task的Process方法 bool ExecuteTaskOnReferenceLine( const common::TrajectoryPoint& planning_start_point, Frame* frame);
bool ExecuteTaskOnReferenceLineForOnlineLearning( const common::TrajectoryPoint& planning_start_point, Frame* frame);
bool ExecuteTaskOnOpenSpace(Frame* frame);
protected: // task的map,key是TaskType,value是指向Task的指针 std::map<TaskConfig::TaskType, std::unique_ptr<Task>> tasks_; // 保存Task列表 std::vector<Task*> task_list_; // stage 配置 ScenarioConfig::StageConfig config_;...};


路径边界决策数据

PathBoundsDecider类主要的输入、输出、数据结构和变量设置。


(1)输入和输出

输入有两个:Framereference_line_info


  • Frame

Frame中包含的一次规划所需要的所有的数据


// modules/planning/common/frame.hclass Frame { private: static DrivingAction pad_msg_driving_action_; uint32_t sequence_num_ = 0;
/* Local_view是一个结构体,包含了如下信息 // modules/planning/common/local_view.h struct LocalView { std::shared_ptr<prediction::PredictionObstacles> prediction_obstacles; std::shared_ptr<canbus::Chassis> chassis; std::shared_ptr<localization::LocalizationEstimate> localization_estimate; std::shared_ptr<perception::TrafficLightDetection> traffic_light; std::shared_ptr<routing::RoutingResponse> routing; std::shared_ptr<relative_map::MapMsg> relative_map; std::shared_ptr<PadMessage> pad_msg; std::shared_ptr<storytelling::Stories> stories; }; */ LocalView local_view_; // 高清地图 const hdmap::HDMap *hdmap_ = nullptr; common::TrajectoryPoint planning_start_point_; // 车辆状态 // modules/common/vehicle_state/proto/vehicle_state.proto common::VehicleState vehicle_state_; // 参考线信息 std::list<ReferenceLineInfo> reference_line_info_;
bool is_near_destination_ = false;
/** * the reference line info that the vehicle finally choose to drive on **/ const ReferenceLineInfo *drive_reference_line_info_ = nullptr;
ThreadSafeIndexedObstacles obstacles_;
std::unordered_map<std::string, const perception::TrafficLight *> traffic_lights_;
// current frame published trajectory ADCTrajectory current_frame_planned_trajectory_;
// current frame path for future possible speed fallback DiscretizedPath current_frame_planned_path_;
const ReferenceLineProvider *reference_line_provider_ = nullptr;
OpenSpaceInfo open_space_info_;
std::vector<routing::LaneWaypoint> future_route_waypoints_;
common::monitor::MonitorLogBuffer monitor_logger_buffer_;};

  • Reference_line_info

reference_line_info包含了有关reference_line的所有数据


// modules/planning/common/reference_line_info.h
class ReferenceLineInfo { ... private: static std::unordered_map<std::string, bool> junction_right_of_way_map_; const common::VehicleState vehicle_state_; // 车辆状态 const common::TrajectoryPoint adc_planning_point_; // TrajectoryPoint定义在modules/common/proto/pnc_point.proto中
/* 参考线,以道路中心线,做过顺滑的一条轨迹,往后80米,往前130米。 class ReferenceLine { ... private: struct SpeedLimit { double start_s = 0.0; double end_s = 0.0; double speed_limit = 0.0; // unit m/s ...}; // This speed limit overrides the lane speed limit std::vector<SpeedLimit> speed_limit_; std::vector<ReferencePoint> reference_points_; // ReferencePoint包含有信息(k, dk, x, y, heading, s, l) hdmap::Path map_path_; uint32_t priority_ = 0; }; */ ReferenceLine reference_line_;
/** * @brief this is the number that measures the goodness of this reference * line. The lower the better. */ // 评价函数,值越低越好 double cost_ = 0.0;
bool is_drivable_ = true; // PathDecision包含了一条路径上的所有obstacle的决策,有两种:lateral(Nudge, Ignore)和longitudinal(Stop, Yield, Follow, Overtake, Ignore) PathDecision path_decision_; // 指针 Obstacle* blocking_obstacle_;
/* path的边界,结果保存在这个变量里。通过**SetCandidatePathBoundaries**方法保存到此变量 // modules/planning/common/path_boundary.h class PathBoundary { ... private: double start_s_ = 0.0; double delta_s_ = 0.0; std::vector<std::pair<double, double>> boundary_; std::string label_ = "regular"; std::string blocking_obstacle_id_ = ""; }; */ std::vector<PathBoundary> candidate_path_boundaries_; // PathData类,包含XY坐标系和SL坐标系的相互转化 std::vector<PathData> candidate_path_data_;
PathData path_data_; PathData fallback_path_data_; SpeedData speed_data_;
DiscretizedTrajectory discretized_trajectory_;
RSSInfo rss_info_;
/** * @brief SL boundary of stitching point (starting point of plan trajectory) * relative to the reference line */ SLBoundary adc_sl_boundary_;... };


输出:

Status PathBoundsDecider::Process( Frame* const frame, ReferenceLineInfo* const reference_line_info)


Process函数定义,最终结果保存到reference_line_info


(2)参数设置

// modules/planning/tasks/deciders/path_bounds_decider/path_bounds_decider.cc// s方向的距离constexpr double kPathBoundsDeciderHorizon = 100.0;// s方向的间隔constexpr double kPathBoundsDeciderResolution = 0.5;
// Lane宽度constexpr double kDefaultLaneWidth = 5.0;// Road的道路constexpr double kDefaultRoadWidth = 20.0;
// TODO(all): Update extra tail point base on vehicle speed.constexpr int kNumExtraTailBoundPoint = 20;constexpr double kPulloverLonSearchCoeff = 1.5;constexpr double kPulloverLatSearchCoeff = 1.25;


(3)数据结构

// modules/planning/tasks/deciders/path_bounds_decider/path_bounds_decider.ccnamespace {// PathBoundPoint contains: (s, l_min, l_max). 路径边界点using PathBoundPoint = std::tuple<double, double, double>;// PathBound contains a vector of PathBoundPoints. 路径边界using PathBound = std::vector<PathBoundPoint>;// ObstacleEdge contains: (is_start_s, s, l_min, l_max, obstacle_id). 障碍物的边using ObstacleEdge = std::tuple<int, double, double, double, std::string>;} // namespace


下图是路径边界决策的流程图:



Process方法中,分四种场景对路径边界进行计算,按照处理的顺序分别是:Fallback、Pull-over、Lane-change、Regular。其中Regular场景根据是否借道又分为:LEFT_BORROWNO_BORROWRIGHT_BORROW


Fallback场景的Path bounds一定会生成,另外三种看情况,都是需要if判断。



1、Fallback



Fallback场景生成过程如上图所示。Fallback只考虑ADC信息和静态道路信息,主要调用两个函数:


InitPathBoundary

bool PathBoundsDecider::InitPathBoundary( ... // Starting from ADC's current position, increment until the horizon, and // set lateral bounds to be infinite at every spot. // 从adc当前位置开始,以0.5m为间隔取点,直到终点,将 [左, 右] 边界设置为double的 [lowerst, max] for (double curr_s = adc_frenet_s_; curr_s < std::fmin(adc_frenet_s_ + std::fmax(kPathBoundsDeciderHorizon, reference_line_info.GetCruiseSpeed() * FLAGS_trajectory_time_length), reference_line.Length()); curr_s += kPathBoundsDeciderResolution) { path_bound->emplace_back(curr_s, std::numeric_limits<double>::lowest(), std::numeric_limits<double>::max()); }...}

GetBoundaryFromLanesAndADC

// TODO(jiacheng): this function is to be retired soon.bool PathBoundsDecider::GetBoundaryFromLanesAndADC( ... for (size_t i = 0; i < path_bound->size(); ++i) { double curr_s = std::get<0>((*path_bound)[i]); // 1. Get the current lane width at current point.获取当前点车道的宽度 if (!reference_line.GetLaneWidth(curr_s, &curr_lane_left_width, &curr_lane_right_width)) { AWARN << "Failed to get lane width at s = " << curr_s; curr_lane_left_width = past_lane_left_width; curr_lane_right_width = past_lane_right_width; } else {...}
// 2. Get the neighbor lane widths at the current point.获取当前点相邻车道的宽度 double curr_neighbor_lane_width = 0.0; if (CheckLaneBoundaryType(reference_line_info, curr_s, lane_borrow_info)) { hdmap::Id neighbor_lane_id; if (lane_borrow_info == LaneBorrowInfo::LEFT_BORROW) { // 借左车道 ... } else if (lane_borrow_info == LaneBorrowInfo::RIGHT_BORROW) { // 借右车道 ... } }
// 3. 根据道路宽度,adc的位置和速度计算合适的边界。 static constexpr double kMaxLateralAccelerations = 1.5; double offset_to_map = 0.0; reference_line.GetOffsetToMap(curr_s, &offset_to_map);
double ADC_speed_buffer = (adc_frenet_ld_ > 0 ? 1.0 : -1.0) * adc_frenet_ld_ * adc_frenet_ld_ / kMaxLateralAccelerations / 2.0; // 向左车道借到,左边界会变成左侧车道左边界 double curr_left_bound_lane = curr_lane_left_width + (lane_borrow_info == LaneBorrowInfo::LEFT_BORROW ? curr_neighbor_lane_width : 0.0); // 和上面类似 double curr_right_bound_lane = -curr_lane_right_width - (lane_borrow_info == LaneBorrowInfo::RIGHT_BORROW ? curr_neighbor_lane_width : 0.0);
double curr_left_bound = 0.0; // 左边界 double curr_right_bound = 0.0; // 右边界 // 计算左边界和右边界 if (config_.path_bounds_decider_config() .is_extend_lane_bounds_to_include_adc() || is_fallback_lanechange) { // extend path bounds to include ADC in fallback or change lane path // bounds. double curr_left_bound_adc = std::fmax(adc_l_to_lane_center_, adc_l_to_lane_center_ + ADC_speed_buffer) + GetBufferBetweenADCCenterAndEdge() + ADC_buffer; curr_left_bound = std::fmax(curr_left_bound_lane, curr_left_bound_adc) - offset_to_map;
double curr_right_bound_adc = std::fmin(adc_l_to_lane_center_, adc_l_to_lane_center_ + ADC_speed_buffer) - GetBufferBetweenADCCenterAndEdge() - ADC_buffer; curr_right_bound = std::fmin(curr_right_bound_lane, curr_right_bound_adc) - offset_to_map; } else { curr_left_bound = curr_left_bound_lane - offset_to_map; curr_right_bound = curr_right_bound_lane - offset_to_map; }
// 4. 更新边界. if (!UpdatePathBoundaryWithBuffer(i, curr_left_bound, curr_right_bound, path_bound, is_left_lane_boundary, is_right_lane_boundary)) { path_blocked_idx = static_cast<int>(i); }... }


2、Pull over



(1)GetBoundaryFromRoads

GetBoundaryFromLanesAndADC不同,GetBoundaryFromRoads函数根据道路信息计算出边界:

1、获取参考线信息

2、对路径上的点,逐点计算

  • 边界

  • 更新


(2)GetBoundaryFromStaticObstacles

根据障碍车调整边界:

1、计算障碍车在Frenet坐标系下的坐标


2、扫描线排序,S方向扫描

  • 只关注在路径边界内的障碍物

  • 只关注在ADC前方的障碍物

  • 将障碍物分解为两个边界,开始和结束


3、映射障碍物ID

  • ADC能从左边通过为True,否则为False


4、逐点检查Path路径上的障碍物

  • 根据新来的障碍物

  • 根据已有的障碍物


(3)SearchPullOverPosition

搜索Pull over位置的过程:

  • 根据pull_over_status.pull_over_type判断是前向搜索(Pull over开始第一个点),还是后向搜索(Pull over末尾后一个点)。

  • 两层循环,外层控制搜索的索引idx,内层控制进一步索引(前向idx+1,后向idx-1)。

  • 根据内外两层循环的索引,判断搜索到的空间是否满足宽度和长度要求,判断是否可以Pull over。


代码如下:

bool PathBoundsDecider::SearchPullOverPosition( const Frame& frame, const ReferenceLineInfo& reference_line_info, const std::vector<std::tuple<double, double, double>>& path_bound, std::tuple<double, double, double, int>* const pull_over_configuration) { const auto& pull_over_status = injector_->planning_context()->planning_status().pull_over();
// 搜索方向,默认前向搜索 bool search_backward = false; // search FORWARD by default
double pull_over_s = 0.0; if (pull_over_status.pull_over_type() == PullOverStatus::EMERGENCY_PULL_OVER) {...}
int idx = 0; if (search_backward) { // 后向搜索,定位pull over末尾的一个点. idx = static_cast<int>(path_bound.size()) - 1; while (idx >= 0 && std::get<0>(path_bound[idx]) > pull_over_s) { --idx; } } else { // 前向搜索,定位emergency pull over开头后的第一个点. while (idx < static_cast<int>(path_bound.size()) && std::get<0>(path_bound[idx]) < pull_over_s) { ++idx; } } // 为pull over搜索到一个可行的位置,主要是确定该区域的宽度和长度 const double pull_over_space_length = kPulloverLonSearchCoeff * VehicleConfigHelper::GetConfig().vehicle_param().length() - FLAGS_obstacle_lon_start_buffer - FLAGS_obstacle_lon_end_buffer; const double pull_over_space_width = (kPulloverLatSearchCoeff - 1.0) * VehicleConfigHelper::GetConfig().vehicle_param().width(); const double adc_half_width = VehicleConfigHelper::GetConfig().vehicle_param().width() / 2.0;
// 2. Find a window that is close to road-edge. /* 这里用了内外两层循环进行搜索,外层循环控制搜索的开始的端点idx。 内层控制另一个端点。根据找到的两个端点,判断区域是否可以pull over */ bool has_a_feasible_window = false; while ((search_backward && idx >= 0 && std::get<0>(path_bound[idx]) - std::get<0>(path_bound.front()) > pull_over_space_length) || (!search_backward && idx < static_cast<int>(path_bound.size()) && std::get<0>(path_bound.back()) - std::get<0>(path_bound[idx]) > pull_over_space_length)) {
while ((search_backward && j >= 0 && std::get<0>(path_bound[idx]) - std::get<0>(path_bound[j]) < pull_over_space_length) || (!search_backward && j < static_cast<int>(path_bound.size()) && std::get<0>(path_bound[j]) - std::get<0>(path_bound[idx]) < pull_over_space_length)) {...} // 找到可行区域,获取停车区域的位置和姿态 if (is_feasible_window) { ... break;} ...} // 外层while...}


3、Lane change



代码流程如下:

Status PathBoundsDecider::GenerateLaneChangePathBound( const ReferenceLineInfo& reference_line_info, std::vector<std::tuple<double, double, double>>* const path_bound) { // 1.初始化,和前面的步骤类似 if (!InitPathBoundary(reference_line_info, path_bound)) {...}

// 2. 根据道路和adc的信息获取一个大致的路径边界 std::string dummy_borrow_lane_type; if (!GetBoundaryFromLanesAndADC(reference_line_info, LaneBorrowInfo::NO_BORROW, 0.1, path_bound, &dummy_borrow_lane_type, true)) {...}
// 3. Remove the S-length of target lane out of the path-bound. GetBoundaryFromLaneChangeForbiddenZone(reference_line_info, path_bound);

// 根据静态障碍物调整边界. if (!GetBoundaryFromStaticObstacles(reference_line_info.path_decision(), path_bound, &blocking_obstacle_id)) {...} ...}

GetBoundaryFromLaneChangeForbiddenZone函数是Lane change重要的函数。运行过程如下:

  • 如果当前位置可以变道,则直接变道;

  • 如果有一个Lane-change的起点,则直接使用它;

  • 逐一检查变道前点的边界,改变边界的值(如果已经过了变道点,则返回)。


void PathBoundsDecider::GetBoundaryFromLaneChangeForbiddenZone( const ReferenceLineInfo& reference_line_info, PathBound* const path_bound) {

// 1.当前位置直接变道。 auto* lane_change_status = injector_->planning_context() ->mutable_planning_status() ->mutable_change_lane(); if (lane_change_status->is_clear_to_change_lane()) { ADEBUG << "Current position is clear to change lane. No need prep s."; lane_change_status->set_exist_lane_change_start_position(false); return; }

// 2.如果已经有一个lane-change的起点,就直接使用它,否则再找一个 double lane_change_start_s = 0.0; if (lane_change_status->exist_lane_change_start_position()) { common::SLPoint point_sl; reference_line.XYToSL(lane_change_status->lane_change_start_position(), &point_sl); lane_change_start_s = point_sl.s(); } else { // TODO(jiacheng): train ML model to learn this. // 设置为adc前方一段距离为变道起始点 lane_change_start_s = FLAGS_lane_change_prepare_length + adc_frenet_s_;
// Update the decided lane_change_start_s into planning-context. // 更新变道起始点的信息 common::SLPoint lane_change_start_sl; lane_change_start_sl.set_s(lane_change_start_s); lane_change_start_sl.set_l(0.0); common::math::Vec2d lane_change_start_xy; reference_line.SLToXY(lane_change_start_sl, &lane_change_start_xy); lane_change_status->set_exist_lane_change_start_position(true); lane_change_status->mutable_lane_change_start_position()->set_x( lane_change_start_xy.x()); lane_change_status->mutable_lane_change_start_position()->set_y( lane_change_start_xy.y()); }
// Remove the target lane out of the path-boundary, up to the decided S. // 逐个检查变道前的点的边界,改变边界的值 for (size_t i = 0; i < path_bound->size(); ++i) { double curr_s = std::get<0>((*path_bound)[i]); if (curr_s > lane_change_start_s) { break; } double curr_lane_left_width = 0.0; double curr_lane_right_width = 0.0; double offset_to_map = 0.0; reference_line.GetOffsetToMap(curr_s, &offset_to_map); if (reference_line.GetLaneWidth(curr_s, &curr_lane_left_width, &curr_lane_right_width)) { double offset_to_lane_center = 0.0; reference_line.GetOffsetToMap(curr_s, &offset_to_lane_center); curr_lane_left_width += offset_to_lane_center; curr_lane_right_width -= offset_to_lane_center; } curr_lane_left_width -= offset_to_map; curr_lane_right_width += offset_to_map;
std::get<1>((*path_bound)[i]) = adc_frenet_l_ > curr_lane_left_width ? curr_lane_left_width + GetBufferBetweenADCCenterAndEdge() : std::get<1>((*path_bound)[i]); std::get<1>((*path_bound)[i]) = std::fmin(std::get<1>((*path_bound)[i]), adc_frenet_l_ - 0.1); std::get<2>((*path_bound)[i]) = adc_frenet_l_ < -curr_lane_right_width ? -curr_lane_right_width - GetBufferBetweenADCCenterAndEdge() : std::get<2>((*path_bound)[i]); std::get<2>((*path_bound)[i]) = std::fmax(std::get<2>((*path_bound)[i]), adc_frenet_l_ + 0.1); }


4、Regular



代码流程如下:
Status PathBoundsDecider::GenerateRegularPathBound( const ReferenceLineInfo& reference_line_info, const LaneBorrowInfo& lane_borrow_info, PathBound* const path_bound, std::string* const blocking_obstacle_id, std::string* const borrow_lane_type) { // 1.初始化边界. if (!InitPathBoundary(reference_line_info, path_bound)) {...}

// 2.根据adc位置和lane信息确定大致的边界 if (!GetBoundaryFromLanesAndADC(reference_line_info, lane_borrow_info, 0.1, path_bound, borrow_lane_type)) {...} // PathBoundsDebugString(*path_bound);
// 3.根据障碍物调整道路边界 if (!GetBoundaryFromStaticObstacles(reference_line_info.path_decision(), path_bound, blocking_obstacle_id)) {...} ...}


流程和上面的几个基本类似,借道有三种类型:

enum class LaneBorrowInfo { LEFT_BORROW, NO_BORROW, RIGHT_BORROW, };



*《路径边界决策

https://github.com/ApolloAuto/apollo/blob/master/docs/technical_documents/tasks/path_bounds_decider_cn.md


以上就是规划模块中的路径边界决策的全部内容,如果大家对Apollo开放平台和套件感兴趣,可以添加『Apollo小哥哥』(微信号:apollo_xzs)为好友,进入技术交流群,跟开发者们一起讨论哦。




©️原创归作者所有,如需转载,请注明出处,另有法律责任。





您可能也对以下帖子感兴趣

文章有问题?点此查看未经处理的缓存