Data Challenge 2: Human Gait Recognition

What's the Challenge?

Current surveillance solutions primarily rely on facial biometric identification using RGB videos or images acquired from CCTV cameras. These systems are significantly compromised when faces are fully or partially covered by masks, hijabs, turbans, or beards, and in cases of visual ambiguity due to changes in illumination, smog, dust, or fog. Additionally, these systems are highly dependent on the point-of-view and do not perform well over long distances. This challenge is prevalent in India and similar regions, especially when surveillance systems need to scale for large populations.

These vulnerabilities undermine the reliability of existing surveillance methods, necessitating the development of a more robust solution. Analyzing a person's identity through body shape and walking posture offers an excellent alternative. This method relies on the fact that each individual has a unique physiological structure, including height, head shape, leg bones, hip extension, muscles, and other factors.

Human gait recognition, using classification of silhouettes obtained from RGB videos and point cloud data from LiDAR sensors, offers a promising alternative. Gait recognition is less affected by obstructions and external visual factors, making it a more robust biometric for surveillance in challenging environments.

Track Description:

Track 1: Multi-view silhouette-based gait recognition
  • Training Dataset: Silhouettes obtained from RGB camera videos captured at 0o to the plane of the person’s walking trajectory.
  • Testing Dataset: Silhouettes obtained from RGB camera videos captured at 90o to the plane of the person’s walking trajectory.
  • Objective: Calculate the F1 score for the classification of human gait using silhouettes. Track 2: Multimodal (silhouette+LiDAR point cloud) gait recognition
  • Training Dataset: Silhouettes obtained from RGB camera videos and point cloud data from LiDAR point cloud videos, captured at 0o to the plane of the person’s walking trajectory.
  • Testing Dataset: Silhouettes obtained from RGB camera videos and point cloud data from LiDAR point cloud videos, captured at 0o to the plane of the person’s walking trajectory. Objective: Calculate the F1 score for the classification of human gait using multimodal (silhouette+LiDAR point cloud) data fusion.

Dataset Information

Dataset Name: IISERB-PS-G

Dataset Description: The dataset contains 2055 subjects of round-trip walking sequences captured by two cameras at 0o and 90oto the plane of the person’s walking trajectory at frame-rate of 30 fps. This dataset will be used for track-1.

Another dataset contains 255 subjects of round-trip walking sequences captured by two cameras at 0o and 90oto the plane of the person’s walking trajectory at frame-rate of 30 fps, in combination with LiDAR point cloud data captured by Velodyne VLP-16 at 0o at frame-rate of 10 fps. This dataset will be used for track-2. The video capturing setup is shown below.

Track1: Multi-view silhouette-based gait recognition

Size:

  • Total size: 2055 (Silhouette)
  • Train: 1600
  • Validation: 400
  • Test: 55

Example of input data:

Fig 1. Person walking in normal condition [a] RGB image captured at 90oto the plane of the person’s walking trajectory, [b] Silhouette derived from (a), [c] RGB image captured at 0oto the plane of the person’s walking trajectory, [d] Silhouette derived from (c).

Track 2: Multimodal (silhouette+LiDAR point cloud) gait recognition

Size:

  • Total size: 255 (Silhouette+Point Cloud)
  • Train: 160
  • Validation: 40
  • Test: 55

Example of input data:

Fig 1. Person walking in normal condition [a] RGB image captured at 90oto the plane of the person’s walking trajectory, [b] Silhouette derived from (a), [c] RGB image captured at 0oto the plane of the person’s walking trajectory, [d] Silhouette derived from (c), [e] Point cloud frame captured at 0oto the plane of the person’s walking trajectory.

Evaluation Criteria

To be eligible for the prizes, participating teams must submit:
  • Well-commented code used for training and evaluating the models, including scripts for preprocessing, training, validation, and testing.
  • A detailed ReadMe file with setup instructions, dependencies, and steps to reproduce the results.
  • A detailed report (format to be provided) explaining the solution, methodologies, and any noteworthy findings.
  • Evaluation Metrics: The primary metric for evaluating submissions will be the F1 Score, which balances precision and recall, making it suitable for assessing classification performance in this context.

Note: Highest performance metric is not the only criterion on which the winners will be decided. The teams will be judged based on overall performance (novelty of model architecture and preprocessing methods used). The final decision on the winners will be made by the judges. Prizes are to be won in 2 different tracks.

To ensure alignment with real-world impact, additional evaluations on winning submissions can be conducted in realistic surveillance scenarios:

  • Deploying the winning models in a live surveillance environment to monitor performance under real-world conditions, including varying illumination, weather, and obstructions.
  • Evaluating the models in different locations and settings (e.g., indoor vs. outdoor) to assess generalization capabilities.

Timelines

  • 1. User registration opens on 23rd Sept 2024.
  • 2. The competition officially launches on September 26, 2024.
  • 3. Last date to submit is November 7, 2024 for the leaderboard.
  • 4. Participants have until November 10, 2024, to submit their final code and reports.
  • 5. Winners will be declared by Nov 20, 2024.

Competition Rules & Regulations

  • 1. Participants can participate either individually or in a team of a maximum of 3 members.
  • 2. Each team should have at least one person from an Indian University or an Indian research lab.
  • 3. Submission of code/implementation details and report is mandatory to be considered for prizes.
  • 4. The organizers will take the final call on the final prize money as well as any modification of the evaluation criteria (if any).
  • 5. The organizers reserve the right to call off the contest if there are not sufficient teams

Number of Awards

Winner Prizes (Each Track)

  • First Prize - INR 5000/-
  • Second Prize - INR 3000/-
  • Third Prize - INR 2000/-

Note: All active participants will get participation certificates.

Registration/Website Link

Please register for this data challenge using the below link and we will send out the dataset and submission link.
Registration: https://forms.gle/jYYWZQLDG8G5bc6u8

Competition Sponsors

1) Netweb TECHNOLOGIES https://netwebindia.com/
2) PawScan.AI https://pawscanai-68de4.web.app/