Skip to content

Latest commit

 

History

History
168 lines (125 loc) · 9.68 KB

README.md

File metadata and controls

168 lines (125 loc) · 9.68 KB
Raw

Human tracking dataset captured by drone.

Introduction

  • We have captured 500 videos by drone in various environment for multi object tracking task. Images were extracted from 30 frame per second video at different rate for training and test purposes. All extracted images are annotated by human with various properties such as occlusion, re-identification, visibility, pose, image truncation.

drawing drawing

Label format

  • Figure below is the example of ground truth file.

drawing

Position 1 2 3 4 5 6 7 8 9 10 11 12
frame_number person_id_index tracking_id_index box_top_left_x box_top_left_y box_width box_height is_valid pose_class occlusion truncated visibility

Description of properties

Key Value type Description
frame_number INT Image frame number
person_id INT Unique number assinged to person in the sequence (i.e., same person in the sequence will have same number regardeless of appear/disapper)
tracking_id INT Unique number assigned to track of person (i.e., when same person disapeared and appeared again in the sequence, this number will be different)
box_top_left_x INT X coordinate of top left corner
box_top_left_y INT Y coordinate of top left corner
box_width INT box width
box_height INT box height
is_valid INT (0/1) 0 : invalud instance (e.g., human in the picture), 1 : valid instance
pose_class INT (0/1/2) 0 : General pose, 1 : Sitting, 2 : Waving hands
occlusion INT (0/1) 0 : occlusion , 1 : no occlusion
truncated INT (0/1) 0 : truncated , 1 : no truncated
visibility FLOAT area of actual visible area in pixel / area of object size in the image predicted by human labeler

Download

  • Download
    • Train images were extracted from video taken from drone at 1 frame per second.
      • This is for model training.
    • Test images were extracted from video taken from drone at 3 frame per second.
      • This is for evaluating the trained model.

Directory Structure

 [root path]/
 └──[test or train]/                  
    └──[seqence_(sequence number)]/
       ├──(sequence_number).txt     <--- ground truth
       └──[images]/
          └──(frame_number).png     <--- image

Visualizer

 python3 visualize_data.py --dir=[root_dir_path] --type=[train or test] --seq=[sequence number( 450 <= test < 500 ,  0 <= train < 450)]
key 'q' 'esc' : exit program
key 'n' '->'  : next frame
key 'p' '<-'  : prev frame

Data Distribution

Dataset

Number of images Description
train 13,500 Dataset used during the learning process
test 4,500 Dataset used only to assess performance
total 18,000

Drone camera angle

Number of images Description
general 14,370 Dataset taken when the angle of the drone camera to the ground is between 10~60 degrees
top 3,630 Dataset taken when the angle of the drone camera to the ground is around 90 degrees
total 18,000

Weather

Number of images Description
sunny 10,320 Dataset taken on a sunny day
cloudy 7,680 Dataset taken on a rainy or cloudy day
total 18,000

Filming location

Number of images Description
general 9,000 Dataset taken from common roads such as side walk and asphalt
green 6,900 Dataset taken in natural environments
play ground 2,100 Dataset taken from outdoor sports venues such as tennis courts and basketball courts
total 18,000

Class Distribution

Human pose class

Number of objects Description
general 134,328 Objects(people) with comman posture, such as walking or standing
sitting 9,598 Objects(people) with sitting posutre
waving hand 542 Object(people) waving hands
total 144,468

Occlusion

Number of objects Description
occluded 49,258 Objects that are obscured by other objects within the image
not occluded 95,210 Objects that appear completely without occlusion within the image
total 144,468

Truncation

Number of objects Description
truncated 10,508 Objects whose part is out of the image
not truncated 133,960 Objects that appear completely without truncation within the image
total 144,468

Data Example

  • Images below are example images from dataset.

drawing drawing

drawing drawing

Acknowledgement

이 데이터는 2021년도 정부(과학기술정보통신부)의 재원으로 정보통신기획평가원의 지원을 받아 수행된 연구의 결과물임 (No.171125972, 인명 구조용 드론을 위한 영상/음성 인지 기술 개발)

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.171125972, Audio-Visual Perception for Autonomous Rescue Drones)