Simple Explanation (ELI5)

It's like having a balcony camera that, thanks to AI, can count how many friends walk by during the day — but at dusk, when the light fades, it struggles and miscounts, just like a kid tallying playmates in bright sunshine but losing track at sunset.

Key Advantages & Current Limits

🟢 Pros: Public webcams are already everywhere, offering a free, city‑wide data source for traffic monitoring at no extra cost; in bright daylight, open‑source models count cars and pedestrians quite accurately. 🔴 Cons: At night or dusk, shadows and low contrast confuse the algorithms; trams, buses and bicycles are detected less well, requiring custom training for each camera angle and vehicle type.

Future Outlook & Takeaways

In the next 3‑5 years, AI models trained on nighttime datasets and using adaptive anchors will count vehicles and people around the clock; cities will be able to adjust traffic lights and bike lanes in real time, boosting road safety and sustainability.

Can public webcams count cars and people in cities?

Article

100%

Imagine watching a city square camera and instantly knowing how many cars, bikes or people are there. A recent study tried exactly that, using public webcams in Brunswick, Germany. Eight open‑source neural networks were tested on 45 photos taken in bright, dusk and night conditions. Researchers manually labeled each image, separating cars, people, bicycles, trucks, trams and buses, then compared those counts with the AI results.

In bright light, the models accurately counted cars and pedestrians, showing that webcams can be a valuable data source for traffic research. At twilight or night, however, accuracy dropped: shadows and low contrast confuse the algorithms. Performance was also weaker for trams, buses, bicycles and trucks, indicating a need to fine‑tune models for each vehicle type and camera view.

The takeaway is clear: public webcams hold a wealth of information, but unlocking their full potential requires specific model training. In the future, smarter AI and richer data could enable real‑time traffic monitoring, safer streets and more sustainable city planning.

Bonus

100%

The neural networks used in the study belong to two‑stage detectors like Faster R-CNN or single‑stage models such as YOLOv5 and SSD. These architectures combine a backbone (often ResNet or CSPDarknet) for feature extraction with a head that predicts bounding boxes and class labels. In bright conditions, high contrast signals facilitate edge separation and improve activation map precision. At dusk, the reduced signal‑to‑noise ratio leads to false positives and missed detections, especially for small objects like bicycles. At night, the dominance of low‑intensity pixels calls for pre‑processing techniques, such as infrared illumination or training on night‑time datasets (Nighttime Object Detection). Moreover, webcam perspective variance (viewing angle, height) affects apparent object scale, requiring specific data augmentation (scaling, rotation) to boost generalization. For vehicles like trams and buses, their elongated shape can confuse anchor‑based detectors; adopting dynamic anchors or anchor‑free methods (e.g., CenterNet) can reduce these errors. Finally, fine‑tuning on a small, labeled image set per webcam (transfer learning) allows the network weights to adapt to local lighting and camera settings, improving recall without sacrificing precision.