这里通过只提取一类跟踪数据集，因此对其更改，如果需要提取多类的直接打开即可，同时呢这个转换为多累跟踪数据集，其中数据的格式为:

<frame_index>,<target_id>,<bbox_left>,<bbox_top>,<bbox_width>,<bbox_height>,<score>,<object_category>,<truncation>,<occlusion>

其中的类别为：

讯享网

ignored regions(0), pedestrian(1),people(2), bicycle(3), car(4), van(5), truck(6), tricycle(7), awning-tricycle(8), bus(9), motor(10),others(11))

# encoding=utf-8 import os import copy import numpy as np import cv2 # import shutil from collections import defaultdict from tqdm import tqdm # ignored regions (0), # ----------------------1~10类是我们需要检测和跟踪的目标 # pedestrian (1), --> 0 # people (2), --> 1 # bicycle (3), --> 2 # car (4), --> 3 # van (5), --> 4 # truck (6), --> 5 # tricycle (7), --> 6 # awning-tricycle (8), --> 7 # bus (9), --> 8 # motor (10), --> 9 # ---------------------- # others (11) # We need 10 classes to detect and tracking cls2id = { 'pedestrian': 0, 'people': 1, 'bicycle': 2, 'car': 3, 'van': 4, 'truck': 5, 'tricycle': 6, 'awning-tricycle': 7, 'bus': 8, 'motor': 9 } id2cls = { 0: 'pedestrian', 1: 'people', 2: 'bicycle', 3: 'car', 4: 'van', 5: 'truck', 6: 'tricycle', 7: 'awning-tricycle', 8: 'bus', 9: 'motor' } def draw_ignore_regions(img, boxes): """ 输入图片ignore regions涂黑 :param img: opencv(numpy array): H×W×C :param boxes: a list of boxes: left(box[0]), top(box[1]), width(box[2]), height(box[3]) :return: """ if img is None: print('[Err]: Input image is none!') return -1 for box in boxes: box = list(map(lambda x: int(x + 0.5), box)) # 四舍五入 img[box[1]: box[1] + box[3], box[0]: box[0] + box[2]] = [0, 0, 0] return img def gen_dot_train_file(data_root, rel_path, out_root, f_name='detrac.train'): """ To generate the dot train file :param data_root: :param rel_path: :param out_root: :param f_name: :return: """ if not (os.path.isdir(data_root) and os.path.isdir(out_root)): print('[Err]: invalid root') return out_f_path = out_root + '/' + f_name cnt = 0 with open(out_f_path, 'w') as f: root = data_root + rel_path seqs = [x for x in os.listdir(root)] seqs.sort() # seqs = sorted(seqs, key=lambda x: int(x.split('_')[-1])) for seq in tqdm(seqs): img_dir = root + '/' + seq # + '/img1' img_list = [x for x in os.listdir(img_dir)] img_list.sort() for img in img_list: if img.endswith('.jpg'): img_path = img_dir + '/' + img if os.path.isfile(img_path): item = img_path.replace(data_root + '/', '') # print(item) f.write(item + '\n') cnt += 1 print('Total {:d} images for training'.format(cnt)) def gen_track_dataset(src_root, dst_root, viz_root=None): """ :param src_root: :param dst_root: :param viz_root: :return: """ if not os.path.isdir(src_root): print('[Err]: invalid sr dir.') return if not os.path.isdir(dst_root): os.makedirs(dst_root) dst_img_root = dst_root + '/images' dst_txt_root = dst_root + '/labels_with_ids' if not os.path.isdir(dst_img_root): os.makedirs(dst_img_root) if not os.path.isdir(dst_txt_root): os.makedirs(dst_txt_root) # 记录每一个序列开始的id, 初始化为0 # track_start_id = 0 track_start_id_dict = defaultdict(int) # 所有类别start id都从0开始 for cls_id in id2cls.keys(): track_start_id_dict[cls_id] = 0 # 记录总的帧数 frame_cnt = 0 seq_names = [x for x in os.listdir(src_root + '/sequences')] seq_names.sort() # 遍历每一个视频序列 for seq in tqdm(seq_names): print('Processing {}:'.format(seq)) seq_img_dir = src_root + '/sequences/' + seq seq_txt_f_path = src_root + '/annotations/' + seq + '.txt' if not (os.path.isdir(seq_img_dir) and os.path.isfile(seq_txt_f_path)): print('[Warning]: invalid src img dir or invalid annotations file(txt).') continue # 创建目标子目录(图片目录和标签目录) dst_seq_img_dir = dst_img_root + '/' + seq if not os.path.isdir(dst_seq_img_dir): os.makedirs(dst_seq_img_dir) dst_seq_txt_dir = dst_txt_root + '/' + seq if not os.path.isdir(dst_seq_txt_dir): os.makedirs(dst_seq_txt_dir) # 记录该视频seq的最大track_id # seq_max_tar_id = 0 seq_max_tra_id_dict = defaultdict(int) for k in id2cls.keys(): seq_max_tra_id_dict[k] = 0 # 视频序列 seq_frame_names = os.listdir(seq_img_dir) seq_frame_names.sort() # 将该序列的标签文件读入二维数组 with open(seq_txt_f_path, 'r', encoding='utf-8') as f_r: label_lines = f_r.readlines() label_n_lines = len(label_lines) seq_label_array = np.zeros((label_n_lines, 10), np.int32) # 解析该视频序列的每一帧 for line_i, line in enumerate(label_lines): line = [int(x) for x in line.strip().split(',')] seq_label_array[line_i] = line # 记录该视频序列每一帧的ignore_regions和检测/跟踪目标 # 这里需要注意的是，数据集的格式为： # <frame_index>,<target_id>,<bbox_left>,<bbox_top>,<bbox_width>,<bbox_height>,<score>,<object_category>,<truncation>,<occlusion> # 数据集的类别为： # ignored regions(0), pedestrian(1),people(2), bicycle(3), car(4), van(5), truck(6), tricycle(7), awning-tricycle(8), bus(9), motor(10),others(11)) # 从中我们可以看到类别为0时是忽略的，对应的box就是忽略的box，因此分两部分处理，把忽略的和真正的类别进行区分处理 # 标签中第七位为类别位, 即<object_category> seq_ignore_box_label = seq_label_array[seq_label_array[:, 7] == 0] seq_obj_boxes = seq_label_array[(seq_label_array[:, 7] > 0) & (seq_label_array[:, 7] < 11)] # np条件索引 seq_ignore_box_dict = defaultdict(list) seq_objs_label_dict = defaultdict(list) # 把忽略的标签提取出来， 这里是针对这个视频的所有帧，同时使用帧作为键值 for label in seq_ignore_box_label: # key: frame_id(start from 1) seq_ignore_box_dict[label[0]].append(label[2:6]) # 把真正的标签提取出来，这里是针对这个视频的所有帧 for label in seq_obj_boxes: # key: frame_id(start from 1) seq_objs_label_dict[label[0]].append(label) # 为此seq维护一个dict记录每个class对对应的target id seq_cls_target_ids_dict = defaultdict(list) tmp_ids_dict = defaultdict(set) for fr_id in seq_objs_label_dict.keys(): # 处理每一帧，键值为帧 fr_labels = seq_objs_label_dict[fr_id] # 遍历当前帧的标签，获取每个目标的类别cls_id、跟踪的id即target_id，然后使用类别作为键值，跟踪的id为值进行字典保存 # 即保存的是，同一个类别的情况下多个目标的id，既可以遍历其轨迹 for label in fr_labels: cls_id = label[7] - 1 target_id = label[1] # seq_cls_target_ids_dict[cls_id].append(target_id) # key: cls_id tmp_ids_dict[cls_id].add(target_id) # for cls_id in tmp_ids_dict.keys(): track_ids = tmp_ids_dict[cls_id] # track_ids = set(track_ids) track_ids = list(track_ids) track_ids.sort() seq_cls_target_ids_dict[cls_id] = track_ids # track_ids = tmp_ids_dict[9] # track_ids = list(track_ids) # track_ids.sort() # seq_cls_target_ids_dict[0] = track_ids # 更新max_track_id for k, v in seq_cls_target_ids_dict.items(): seq_max_tra_id_dict[k] = len(v) # print("{} max track id: {:d}, start id: {:d}" # .format(id2cls[9], seq_max_tra_id_dict[0], track_start_id_dict[0])) for k in id2cls.keys(): print("{} max track id: {:d}, start id: {:d}" .format(id2cls[k], seq_max_tra_id_dict[k], track_start_id_dict[k])) # 读取每一帧 for fr_id in seq_objs_label_dict.keys(): # ----- fr_labels = seq_objs_label_dict[fr_id] # ----- 读取图片宽高 fr_name = '{:07d}.jpg'.format(fr_id) fr_path = seq_img_dir + '/' + fr_name if not os.path.isfile(fr_path): print('[Err]: invalid image file {}.'.format(fr_path)) continue # H×W×C: BGR img = cv2.imread(fr_path, cv2.IMREAD_COLOR) if img is None: print('[Err]: empty image.') continue H, W, C = img.shape # ----- 绘制ignore regions draw_ignore_regions(img, seq_ignore_box_dict[fr_id]) # ----- 拷贝image到目标目录 dst_img_path = dst_seq_img_dir + '/' + fr_name if not os.path.isfile(dst_img_path): cv2.imwrite(dst_img_path, img) # 将绘制过ignore region的图片存入目标子目录 # print('{} saved to {}'.format(fr_path, dst_seq_img_dir)) # ----- 如果可视化目录不为空, 进行可视化计算 if not (viz_root is None): # 图片可视化目录和路径 viz_dir = viz_root + '/' + seq if not os.path.isdir(viz_dir): os.makedirs(viz_dir) viz_path = viz_dir + '/' + fr_name # 深拷贝一份img数据作为可视化输出 img_viz = copy.deepcopy(img) # ----- 生成label文件(txt) # 记录该帧的每一行label_str(对应一个检测or跟踪目标) fr_label_strs = [] for label in fr_labels: # cls_id and cls_name obj_type = label[7] assert 0 < obj_type < 11 cls_id = obj_type - 1 # 从0开始 # cls_name = id2cls[cls_id] # 这里只获取电动自行车的轨迹标签数据，如果想要其他的，把这个打开即可 if cls_id==9: target_id = label[1] # 记录该target(object)的track id(从1开始: 标签中从0开始) track_id = seq_cls_target_ids_dict[cls_id].index(target_id) + 1 + track_start_id_dict[cls_id] # track_id = target_id bbox_left = label[2] bbox_top = label[3] bbox_width = label[4] bbox_height = label[5] score = label[6] truncation = label[ 8] # no truncation = 0 (truncation ratio 0%), and partial truncation = 1 (truncation ratio 1% °´ 50%)) occlusion = label[9] if occlusion > 1: # heavy occlusion = 2 (occlusion ratio 50% ~ 100%)). # print('[Warning]: skip the bbox because of heavy occlusion') continue # ----- 绘制该label(一个label是一张图的一个检测/跟踪目标): 在归一化之前 if not (viz_root is None): # 如果可视化目录不为空 # 为target绘制bbox pt_1 = (int(bbox_left + 0.5), int(bbox_top + 0.5)) pt_2 = (int(bbox_left + bbox_width), int(bbox_top + bbox_height)) cv2.rectangle(img_viz, pt_1, pt_2, (0, 255, 0), 2) # 绘制类别文字 cls_str = id2cls[cls_id] veh_type_str_size = cv2.getTextSize(cls_str, cv2.FONT_HERSHEY_PLAIN, 1.3, 1)[0] cv2.putText(img_viz, cls_str, (pt_1[0], pt_1[1] + veh_type_str_size[1] + 8), cv2.FONT_HERSHEY_PLAIN, 1.3, [225, 255, 255], 1) # 绘制track id tr_id_str = str(track_id) tr_id_str_size = cv2.getTextSize(tr_id_str, cv2.FONT_HERSHEY_PLAIN, 1.3, 1)[0] cv2.putText(img_viz, tr_id_str, (pt_1[0], pt_1[1] + veh_type_str_size[1] + tr_id_str_size[1] + 8), cv2.FONT_HERSHEY_PLAIN, 1.3, [225, 255, 255], 1) # 计算bbox中心点坐标 bbox_center_x = bbox_left + bbox_width * 0.5 bbox_center_y = bbox_top + bbox_height * 0.5 # 对bbox进行归一化([0.0, 1.0]) bbox_center_x /= W bbox_center_y /= H bbox_width /= W bbox_height /= H # 组织label的内容, 每帧label生成完成才输出 # class_id, track_id, bbox_center_x, box_center_y, bbox_width, bbox_height # label_str = '{:d} {:d} {:.6f} {:.6f} {:.6f} {:.6f}\n'.format( # cls_id, # track_id, # bbox_center_x, # bbox_center_y, # bbox_width, # bbox_height) label_str = '{:d} {:d} {:.6f} {:.6f} {:.6f} {:.6f}\n'.format( 0, track_id, bbox_center_x, bbox_center_y, bbox_width, bbox_height) fr_label_strs.append(label_str) # ----- 输出可视化结果 if not (viz_root is None): # 如果可视化目录不为空 cv2.imwrite(viz_path, img_viz) # ----- 这一帧的targets解析结束才输出一次 # 输出该图片的所有label(一个label对应一个bbox) label_f_path = dst_seq_txt_dir + '/' + fr_name.replace('.jpg', '.txt') with open(label_f_path, 'w', encoding='utf-8') as f: for label_str in fr_label_strs: f.write(label_str) # print('{} written.'.format(label_f_path)) frame_cnt += 1 # 处理完成该视频seq, 更新track_start_id for cls_id in id2cls.keys(): track_start_id_dict[cls_id] += seq_max_tra_id_dict[cls_id] print('Processing seq {} done.\n'.format(seq)) print('Total {:d} frames'.format(frame_cnt)) if __name__ == '__main__': # gen_track_dataset(src_root='E:/MOTdata/VisDrone2020/VisDrone2019-MOT-train', # dst_root='E:/MOTdata/VisDrone2020/VisDrone2019', # viz_root='E:/MOTdata/VisDrone2020/viz_result') gen_track_dataset(src_root='E:/MOTdata/VisDrone2020/VisDrone2019-MOT-val', dst_root='E:/MOTdata/VisDrone2020/VisDrone2019-val', viz_root='E:/MOTdata/VisDrone2020/viz_result_val')

转换完成以后，进行生成.train文件：

讯享网''' _oo0oo_ oo 88" . "88 (| -_- |) 0\ = /0 ___/`---'\___ .' \\| |// '. / \\||| : |||// \ / _||||| -:- |||||- \ | | \\\ - /// | | | \_| ''\---/'' |_/ | \ .-\__ '-' ___/-. / ___'. .' /--.--\ `. .'___ ."" '< `.___\_<|>_/___.' >' "". | | : `- \`.;`\ _ /`;.`/ - ` : | | \ \ `_. \_ __\ /__ _/ .-` / / =====`-.____`.___ \_____/___.-`___.-'===== `=---=' ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 佛祖保佑 永不宕机 永无BUG 佛曰: 写字楼里写字间，写字间里程序员； 程序人员写程序，又拿程序换酒钱。 酒醒只在网上坐，酒醉还来网下眠； 酒醉酒醒日复日，网上网下年复年。 但愿老死电脑间，不愿鞠躬老板前； 奔驰宝马贵者趣，公交自行程序员。 别人笑我忒疯癫，我笑自己命太贱； 不见满街漂亮妹，哪个归得程序员？ Description: 学习使用 Version: 1.0 Author: 赵守风 Email: @.com Date: 2021-05-28 09:45:42 LastEditors: zsf FilePath: \src\gen_data_path copy.py LastEditTime: 2021-05-28 15:23:35 ''' import os import glob import _init_paths from tqdm import tqdm def gen_caltech_path(root_path): label_path = 'labels_with_ids' real_path = os.path.join(root_path, label_path) image_path = real_path.replace('labels_with_ids', 'images') images_exist = sorted(glob.glob(image_path + '/*.png')) with open('E:/MOTdata/VisDrone2020/VisDrone2019.all', 'w') as f: labels = sorted(glob.glob(real_path + '/*.txt')) for label in labels: image = label.replace('labels_with_ids', 'images').replace('.txt', '.png') if image in images_exist: print(image[22:], file=f) f.close() def gen_dot_train_file(data_root, rel_path, out_root, f_name='VisDrone2019.train'): """ To generate the dot train file :param data_root: :param rel_path: :param out_root: :param f_name: :return: """ if not (os.path.isdir(data_root) and os.path.isdir(out_root)): print('[Err]: invalid root') return out_f_path = out_root + '/' + f_name cnt = 0 with open(out_f_path, 'w') as f: root = data_root +'/'+ rel_path seqs = [x for x in os.listdir(root)] seqs.sort() # seqs = sorted(seqs, key=lambda x: int(x.split('_')[-1])) for seq in tqdm(seqs): img_dir = root + '/' + seq # + '/img1' img_list = [x for x in os.listdir(img_dir)] img_list.sort() for img in img_list: if img.endswith('.jpg'): img_path = img_dir + '/' + img if os.path.isfile(img_path): item = img_path.replace(data_root + '/', '') # print(item) f.write(item + '\n') cnt += 1 print('Total {:d} images for training'.format(cnt)) def clean_train_set(img_root, label_root): """ 清理图片个数与标签文件个数不匹配的问题 :param img_root: :param label_root: :return: """ if not (os.path.isdir(img_root) and os.path.isdir(label_root)): print('[Err]: incalid root!') return img_dirs = [img_root + '/' + x for x in os.listdir(img_root)] label_dirs = [label_root + '/' + x for x in os.listdir(label_root)] assert (len(img_dirs) == len(label_dirs)) # 按视频seq名称排序 img_dirs.sort() label_dirs.sort() for img_dir, label_dir in tqdm(zip(img_dirs, label_dirs)): # 一个couple一个couple的检查 for img_name in os.listdir(img_dir): # print(img_name) txt_name = img_name.replace('.jpg', '.txt') txt_path = label_dir + txt_name img_path = img_dir + img_name if os.path.isfile(img_path) and os.path.isfile(txt_path): continue # 两者同时存在, 无需处理 elif os.path.isfile(img_path) and (not os.path.isfile(txt_path)): os.remove(img_path) print('{} removed.'.format(img_path)) elif os.path.isfile(txt_path) and (not os.path.isfile(img_path)): os.remove(txt_path) print('{} removed.'.format(txt_path)) if __name__ == '__main__': #先清理数据不匹配问题 clean_train_set('E:\MOTdata\VisDrone2020\VisDrone2019\images','E:\MOTdata\VisDrone2020\VisDrone2019\labels_with_ids') # 生成VisDrone2020.train文件 root = 'E:/MOTdata/VisDrone2020/VisDrone2019' gen_dot_train_file(root, 'images', 'E:/MOTdata/VisDrone2020') # gen_caltech_path(root)