Master Dynamic Road Condition Analysis with Car Video – AMAP-TECH Competition Overview

Background

Gaode Map provides massive positioning and navigation services, and the accuracy of its road‑condition status (smooth, slow, congested) directly affects user travel decisions. Traditional road‑condition estimation relies on vehicle trajectory data, which is unreliable on low‑traffic or abnormal roads. In‑vehicle video captures richer information such as vehicle count, road width, and openness, enabling more accurate road‑condition assessment.

Problem Description and Data

The competition requires participants to use computer‑vision and AI algorithms to infer road‑condition status from video frames, improving the accuracy of Gaode Map’s traffic information.

Terminology

Road condition: smooth, slow, congested (displayed as green, yellow, red).

Reference frame: the frame in a sequence whose road‑condition label is considered the ground truth.

Task Definition

Input: a sequence of 3‑5 images with GPS timestamps, one of which is the reference frame.

Output: the road‑condition label (smooth, slow, congested) for the sequence, based on the reference frame.

When the sequence contains inconsistent conditions, the reference frame determines the final label.

Data Description

The dataset is split into a pre‑competition set (released for the preliminary round) and a final set (released after the preliminaries). The pre‑competition set contains 1,500 sequences (~7,000 images) for training and 600 sequences (~2,800 images) for testing. Road‑condition distribution is approximately 70% smooth, 10% slow, 20% congested. Labels are based mainly on the reference frame, with additional frame information used when vehicle density is high.

Data Format

Each sequence is stored in a folder containing up to five frames. An accompanying JSON annotation provides:

Reference frame file name.

Road‑condition status (0: smooth, 1: slow, 2: congested, -1: unknown for test set).

GPS timestamps for each frame (in seconds).

{
    "annotations": [
        {
            "id": "000001",
            "key_frame": "2.jpg",
            "status": 0,
            "frames": [
                {"frame_name": "1.jpg", "gps_time": 1552806921},
                {"frame_name": "2.jpg", "gps_time": 1552806926},
                {"frame_name": "3.jpg", "gps_time": 1552806931},
                {"frame_name": "4.jpg", "gps_time": 1552806936}
            ]
        },
        {
            "id": "000002",
            "key_frame": "3.jpg",
            "status": 2,
            "frames": [
                {"frame_name": "1.jpg", "gps_time": 1555300555},
                {"frame_name": "2.jpg", "gps_time": 1555300560},
                {"frame_name": "3.jpg", "gps_time": 1555300565},
                {"frame_name": "4.jpg", "gps_time": 1555300570},
                {"frame_name": "5.jpg", "gps_time": 1555300580}
            ]
        }
    ]
}

Evaluation

Road‑condition classification (smooth/slow/congested) is evaluated using a weighted F1‑Score. Weights are 0.2 for smooth, 0.2 for slow, and 0.6 for congested.

Expert Judges

Judging panel includes professors and researchers from Peking University, Chinese Academy of Sciences, and senior experts from Alibaba Gaode Map.

Schedule and Participants

Registration: July 8 – August 28

Preliminary round: July 8 – August 31

Final round: September 4 – October 13

Final competition: late October (date TBD)

The dataset becomes publicly downloadable on July 8. Submissions of JSON results can be uploaded online after July 20 10:00 AM for evaluation.

Prizes

Champion: ¥60,000 + certificate

Runner‑up: ¥40,000 + certificate

Third place: ¥20,000 + certificate

Two Excellence awards: ¥10,000 each + certificate

Top‑10 teams in the final round may receive a fast‑track recruitment channel at Alibaba Gaode Map.