YOLOv8s fine-tuned for elevated-camera traffic scenes (VisDrone)

A YOLOv8s detector fine-tuned on VisDrone2019-DET and remapped to six traffic classes. It is the optional detector of the Traffic-Monitor project, a pipeline that reports jams, crashes, stopped vehicles and pedestrian conflicts from video.

Short version: much better than the stock COCO model on drone / elevated views of vehicles, but it learned from drone footage, so check it on your own footage — on one dashcam clip it found fewer vehicles than the stock model.

Classes

The classes keep the COCO ids, so the model is a drop-in replacement wherever yolov8s.pt is used with a class filter:

Ids 4 and 6 (unused4, unused6) exist only to keep the other ids aligned with COCO and are never predicted.

Results

Evaluated on VisDrone2019-DET validation (548 images, 38,759 boxes), imgsz 960. The same split was used to pick the best epoch, so these numbers are slightly optimistic; no test-dev evaluation was run.

Per class, mAP50-95: car 0.598 · bus 0.462 · truck 0.310 · motorcycle 0.259 · person 0.256 · bicycle 0.113.

The matrix is computed at Ultralytics' validation confidence of 0.001, which inflates cross-class confusion and the "background" column. It still shows the real weak spots: motorcycles, bicycles and people are mixed up with each other (only about a third of motorcycles and bicycles are labelled correctly) and 40% of trucks are called cars.

Training log

Validation metrics during training (VisDrone validation set, full per-epoch log in training/results.csv):

Improvement was steady and had flattened by epoch 25-26; Ultralytics keeps the best epoch (26 by mAP50-95) as the shipped weights. The final re-validation above (0.553 / 0.333) differs slightly from the in-training log because it re-runs validation with the maximum detections raised to match the densest image (up to 902 objects).

Demo videos

demos/ has an annotated demo of the Traffic-Monitor pipeline on a scripted synthetic scene, with its event log. It was made with the stock YOLOv8s detector, not with these weights, and shows the decision rules rather than this model. Details and caveats: demos/README.md.

Limitations — please read

• Domain: trained only on drone footage. In a small spot check on 24 frames, it found about twice as many vehicles as the stock model on elevated or fixed cameras, but fewer on a dashcam clip (1.8 vs 5.2 per frame). Detection counts are not accuracy; that was not measured on real traffic-camera footage.
• Weak on bicycles, trucks, buses and riders (see above). Class merges (pedestrian + people, car + van) lose information.
• Small objects at low resolution are still missed; training and evaluation used imgsz 960.
• Not intended for safety-critical use.

Usage

from ultralytics import YOLO
model = YOLO("yolov8s_traffic_best.pt")
results = model.predict("road.jpg", imgsz=960, conf=0.4, classes=[0, 1, 2, 3, 5, 7])

With the Traffic-Monitor pipeline:

python -m traffic_watch --source clip.mp4 --weights yolov8s_traffic_best.pt --imgsz 960

Training

Base yolov8s.pt; 30 epochs, imgsz 960, batch 8, AdamW (Ultralytics auto), cosine LR, mosaic off for the last 3 epochs, AMP, one Tesla T4 (1.75 h), Ultralytics 8.4.171. Training set: VisDrone2019-DET train (6,471 images, 343,205 boxes after remapping). Reproduce with notebooks/01_yolo_finetune_traffic.ipynb in the GitHub repository.

License and data

• Weights: AGPL-3.0, inherited from Ultralytics YOLOv8.
• Training data: VisDrone (Tianjin University, AISKYEYE). It ships without a licence file and is intended for research use; check its terms before any commercial use. Cite: Du et al., "Detection and Tracking Meet Drones Challenge", IEEE TPAMI 2022.

Files in this repository