import os import math from itertools import combinations import numpy as np from scipy.spatial import KDTree import logging import pandas as pd from datetime import datetime, timedelta class GPSFilter: """过滤密集点及孤立点""" def __init__(self, output_dir): self.logger = logging.getLogger('UAV_Preprocess.GPSFilter') @staticmethod def _haversine(lat1, lon1, lat2, lon2): """计算两点之间的地理距离（单位：米）""" R = 6371000 # 地球平均半径，单位：米 phi1, phi2 = math.radians(lat1), math.radians(lat2) delta_phi = math.radians(lat2 - lat1) delta_lambda = math.radians(lon2 - lon1) a = math.sin(delta_phi / 2) ** 2 + math.cos(phi1) * \ math.cos(phi2) * math.sin(delta_lambda / 2) ** 2 c = 2 * math.atan2(math.sqrt(a), math.sqrt(1 - a)) return R * c @staticmethod def _assign_to_grid(lat, lon, grid_size, min_lat, min_lon): """根据经纬度和网格大小，将点分配到网格""" grid_x = int((lat - min_lat) // grid_size) grid_y = int((lon - min_lon) // grid_size) return grid_x, grid_y def _get_distances(self, points_df, grid_size): """读取图片 GPS 坐标，计算点对之间的距离并排序""" # 确定经纬度范围 min_lat, max_lat = points_df['lat'].min(), points_df['lat'].max() min_lon, max_lon = points_df['lon'].min(), points_df['lon'].max() self.logger.info( f"经纬度范围：纬度[{min_lat:.6f}, {max_lat:.6f}]，纬度范围[{max_lat-min_lat:.6f}]，" f"经度[{min_lon:.6f}, {max_lon:.6f}]，经度范围[{max_lon-min_lon:.6f}]") # 分配到网格 grid_map = {} for _, row in points_df.iterrows(): grid = self._assign_to_grid( row['lat'], row['lon'], grid_size, min_lat, min_lon) if grid not in grid_map: grid_map[grid] = [] grid_map[grid].append((row['file'], row['lat'], row['lon'])) self.logger.info(f"图像点已分配到 {len(grid_map)} 个网格中") # 在每个网格中计算两两距离并排序 sorted_distances = {} for grid, images in grid_map.items(): distances = [] for (img1, lat1, lon1), (img2, lat2, lon2) in combinations(images, 2): dist = self._haversine(lat1, lon1, lat2, lon2) distances.append((img1, img2, dist)) distances.sort(key=lambda x: x[2]) # 按距离升序排序 sorted_distances[grid] = distances self.logger.debug(f"网格 {grid} 中计算了 {len(distances)} 个距离对") return sorted_distances def _group_by_time(self, points_df: pd.DataFrame, time_threshold: timedelta) -> list: """根据拍摄时间分组图片 如果相邻两张图片的拍摄时间差超过5分钟，则进行切分 Args: points_df: 包含图片信息的DataFrame，必须包含'file'和'date'列 time_threshold: 时间间隔阈值，默认5分钟 Returns: list: 每个元素为时间组内的点数据 """ if 'date' not in points_df.columns: self.logger.error("数据中缺少date列") return [points_df] # 将date为空的行单独作为一组 null_date_group = points_df[points_df['date'].isna()] valid_date_points = points_df[points_df['date'].notna()] if not null_date_group.empty: self.logger.info(f"发现 {len(null_date_group)} 个无时间戳的点，将作为单独分组") if valid_date_points.empty: self.logger.warning("没有有效的时间戳数据") return [null_date_group] if not null_date_group.empty else [] # 按时间排序 valid_date_points = valid_date_points.sort_values('date') self.logger.info( f"有效时间范围: {valid_date_points['date'].min()} 到 {valid_date_points['date'].max()}") # 计算时间差 time_diffs = valid_date_points['date'].diff() # 找到时间差超过阈值的位置 time_groups = [] current_group_start = 0 for idx, time_diff in enumerate(time_diffs): if time_diff and time_diff > time_threshold: # 添加当前组 current_group = valid_date_points.iloc[current_group_start:idx] time_groups.append(current_group) # 记录断点信息 break_time = valid_date_points.iloc[idx]['date'] group_start_time = current_group.iloc[0]['date'] group_end_time = current_group.iloc[-1]['date'] self.logger.info( f"时间组 {len(time_groups)}: {len(current_group)} 个点, " f"时间范围 [{group_start_time} - {group_end_time}]" ) self.logger.info( f"在时间 {break_time} 处发现断点，时间差为 {time_diff}") current_group_start = idx # 添加最后一组 last_group = valid_date_points.iloc[current_group_start:] if not last_group.empty: time_groups.append(last_group) self.logger.info( f"时间组 {len(time_groups)}: {len(last_group)} 个点, " f"时间范围 [{last_group.iloc[0]['date']} - {last_group.iloc[-1]['date']}]" ) # 如果有空时间戳的点，将其作为最后一组 if not null_date_group.empty: time_groups.append(null_date_group) self.logger.info(f"添加无时间戳组: {len(null_date_group)} 个点") self.logger.info(f"共分为 {len(time_groups)} 个时间组") return time_groups def filter_dense_points(self, points_df, grid_size=0.001, distance_threshold=13, time_threshold=timedelta(minutes=5)): """ 过滤密集点，先按时间分组，再在每个时间组内过滤。 空时间戳的点不进行过滤。 Args: points_df: 点数据 grid_size: 网格大小 distance_threshold: 距离阈值（米） time_interval: 时间间隔（秒） """ self.logger.info(f"开始按时间分组过滤密集点 (网格大小: {grid_size}, " f"距离阈值: {distance_threshold}米, 分组时间间隔: {time_threshold}秒)") # 按时间分组 time_groups = self._group_by_time(points_df, time_threshold) # 存储所有要删除的图片 all_to_del_imgs = [] # 对每个时间组进行密集点过滤 for group_idx, group_points in enumerate(time_groups): # 检查是否为空时间戳组（最后一组） if group_idx == len(time_groups) - 1 and group_points['date'].isna().any(): self.logger.info(f"跳过无时间戳组 (包含 {len(group_points)} 个点)") continue self.logger.info( f"处理时间组 {group_idx + 1} (包含 {len(group_points)} 个点)") # 计算该组内的点间距离 sorted_distances = self._get_distances(group_points, grid_size) group_to_del_imgs = [] # 在每个网格中过滤密集点 for grid, distances in sorted_distances.items(): grid_del_count = 0 while distances: candidate_img1, candidate_img2, dist = distances[0] if dist < distance_threshold: distances.pop(0) # 获取候选图片的其他最短距离 candidate_img1_dist = None candidate_img2_dist = None for distance in distances: if candidate_img1 in distance: candidate_img1_dist = distance[2] break for distance in distances: if candidate_img2 in distance: candidate_img2_dist = distance[2] break # 选择要删除的点 if candidate_img1_dist and candidate_img2_dist: to_del_img = candidate_img1 if candidate_img1_dist < candidate_img2_dist else candidate_img2 group_to_del_imgs.append(to_del_img) grid_del_count += 1 self.logger.debug( f"时间组 {group_idx + 1} 网格 {grid} 删除密集点: {to_del_img} (距离: {dist:.2f}米)") distances = [ d for d in distances if to_del_img not in d] else: break if grid_del_count > 0: self.logger.info( f"时间组 {group_idx + 1} 网格 {grid} 删除了 {grid_del_count} 个密集点") all_to_del_imgs.extend(group_to_del_imgs) self.logger.info( f"时间组 {group_idx + 1} 共删除 {len(group_to_del_imgs)} 个密集点") # 过滤数据 filtered_df = points_df[~points_df['file'].isin(all_to_del_imgs)] self.logger.info( f"密集点过滤完成，共删除 {len(all_to_del_imgs)} 个点，剩余 {len(filtered_df)} 个点") return filtered_df def filter_isolated_points(self, points_df, threshold_distance=0.001, min_neighbors=6): """过滤孤立点""" self.logger.info( f"开始过滤孤立点 (距离阈值: {threshold_distance}, 最小邻居数: {min_neighbors})") coords = points_df[['lat', 'lon']].values kdtree = KDTree(coords) neighbors_count = [len(kdtree.query_ball_point( coord, threshold_distance)) for coord in coords] isolated_points = [] for i, (_, row) in enumerate(points_df.iterrows()): if neighbors_count[i] < min_neighbors: isolated_points.append(row['file']) self.logger.debug( f"删除孤立点: {row['file']} (邻居数: {neighbors_count[i]})") filtered_df = points_df[~points_df['file'].isin(isolated_points)] self.logger.info( f"孤立点过滤完成，共删除 {len(isolated_points)} 个点，剩余 {len(filtered_df)} 个点") return filtered_df