What's the most reliable object detection model for counting vehicles in smart city traffic feeds?

For vehicle counting in smart-city traffic feeds, the most reliable choice in practice is usually:

YOLOv8 / YOLOv10-family detector + tracking

If you want a single model family that is accurate, fast, and easy to deploy, YOLOv8 (or newer YOLOv10, depending on your stack) is often the best starting point for live traffic video.

Why this is a strong choice

• Real-time speed for edge or server deployment
• Good accuracy on cars, buses, trucks, motorcycles
• Strong ecosystem for retraining and deployment
• Works well with multi-object tracking for counting vehicles across frames

Important note: counting is not just detection

For reliable vehicle counts, you generally want:

1. Object detector: detects vehicles in each frame
2. Tracker: keeps identity across frames
   • Common choices: ByteTrack, DeepSORT, BoT-SORT
3. Counting logic: count when an ID crosses a virtual line or enters a zone

Without tracking, simple detection-based counting will usually double count the same vehicle.

Best options by use case

If you need maximum practical reliability for live traffic:

• YOLOv8 + ByteTrack
• Very common production setup
• Good balance of accuracy and throughput

If you need higher accuracy and can afford more compute:

• Faster R-CNN or Cascade R-CNN
• Often more accurate in complex scenes
• Slower, usually less suitable for real-time city cameras

If you need top-end transformer-based detection:

• RT-DETR
• Strong accuracy/speed balance
• Worth considering if you’re evaluating newer models

What matters more than the model

For smart-city traffic feeds, reliability often depends more on:

• Training on local camera data
• Weather/night/rain robustness
• Camera angle and occlusion handling
• Class definitions: car, bus, truck, motorcycle, bike
• Tracking quality
• Calibration of count lines/zones

A model trained on general datasets like COCO may perform well initially, but fine-tuning on your specific traffic cameras usually gives the biggest gain.

Practical recommendation

If you want one answer:

Use YOLOv8 or YOLOv10 fine-tuned on your traffic footage, paired with ByteTrack for counting.

If you want, I can also recommend:

• the best model for edge devices vs GPU servers
• a benchmark comparison of YOLOv8, RT-DETR, Faster R-CNN, and DETR
• a full vehicle counting pipeline architecture