2026年4月施行 道路交通法改正対応
自転車の交通反則通告制度（青切符）の対象となる113種類の違反のうち、 映像解析で検知可能な5種類をAIで自動検出。16歳以上が対象、反則金5,000〜12,000円。
自転車運転中のスマホ操作を検知。YOLO cell phone検出 + 人物-自転車ペアリング。反則金 12,000円。
赤信号での交差点通過を検知。信号ROI色判定 + crossing zone内の移動体検出。反則金 6,000円。
傘を差しながらの自転車運転を検知。umbrella検出 + cyclist bboxオーバーラップ判定。反則金 6,000円。
車道右側の逆走を検知。移動方向角度 vs 期待走行方向で判定。反則金 6,000円。
歩道上の自転車走行を検知。ROIポリゴンで歩道エリアを定義し位置判定。反則金 6,000円。
路上喫煙禁止区域での歩きタバコを検知。タバコ検出モデル + VLM検証。条例違反 過料2,000円。
Architecture
A four-stage pipeline processes each video frame. All thresholds and rules are externalized to a YAML config file.
YOLOv8 detects persons, bicycles, phones, umbrellas, and other objects in each frame.
IoU-based SORT tracker assigns persistent IDs to detected objects across frames.
Rule-based analyzers evaluate spatial relationships, movement patterns, and temporal persistence.
Violations are logged as structured JSON events and overlaid on the output video.
Live Results
Sample detection results from the pipeline. Select a video to view detected events.
Output Format
Each run produces a structured JSON file with all detected violation events.
Loading...
City-Scale Analysis
Geo-referenced violations from multiple camera locations. Click a marker for details.
What's Next
Planned improvements and extensions.
YOLO detection, SORT tracking, rule-based violation analysis, JSON + video output.
Detect bicycles running red lights using traffic signal state recognition.
Detect bicycles traveling against the expected traffic direction.
Detect bicycles on pedestrian sidewalks using ROI-based zone detection.
Replace proxy-object heuristics with keypoint-based hand-to-mouth detection.
Geo-referenced violation heatmaps for city-scale analysis.