Air Quality Prediction & Pollution Hotspot Mapper
Build a pipeline that ingests public air-quality sensor data, forecasts PM2.5/AQI with uncertainty, and maps pollution hotspots — informational environmental...
Build a computer-vision pipeline that classifies camera frames into fog and visibility levels with confidence — a prototype visibility-monitoring system with honest limits and no safety guarantees.

A computer-vision project that watches camera frames — from a webcam, traffic camera, or any video source — and classifies each frame’s atmospheric condition into visibility levels (clear, light fog, dense fog, and so on), producing a visibility estimate with a confidence score and a time-series alert log. It is the atmospheric-visibility counterpart of an image classifier: instead of recognizing objects in a scene, it reasons about how well the scene can be seen, which makes it a genuinely different computer-vision job.
>This is a prototype monitoring aid, not a safety system. Visibility estimates depend on camera quality, lighting, and conditions a model cannot see (night, rain on the lens, sensor noise), so the output is a research-grade indication — never a guarantee of road safety, driving readiness, or accident prevention. It is built to inform humans and to be validated against real conditions before anyone acts on it.
Fog develops faster than forecasts refresh, and it is intensely local: one valley can be in dense fog while the next ridge is clear. Road and weather services would benefit from cheap, camera-based fog detection, but most visibility sensors are specialized instruments that cost too much to deploy broadly. Meanwhile, the cameras already pointing at roads, harbors, and campuses capture exactly the visual signal needed — if a model could turn “how foggy does this frame look” into a usable visibility category. The missing piece is a well-scoped classifier: one that detects atmospheric visibility degradation, reports confidence honestly, and records a time series that a human can review.
The project needs labeled fog/visibility data. Established public options include Foggy Cityscapes and FRIDA, research datasets of foggy road scenes built by simulating or capturing fog over real imagery. For a local MVP, a smaller self-built set can be labeled by hand into coarse classes (clear / light fog / dense fog) from public webcam archives. The data is split into train/validation/test sets with temporal awareness — frames from the same camera should not appear in both training and test.
Each frame is resized and normalized like any classifier input. Because fog is a scene-level property, global statistics matter: preprocessing can include per-frame brightness/histogram features as model inputs alongside the image itself, giving the model explicit cues about contrast loss.
Two complementary feature families feed the model:
Combining both makes the model’s decisions more interpretable and more robust to non-fog scene differences.
Two viable framings, and the project should pick one explicitly:
A sensible intermediate: classification for alerting plus a continuous “fog score” that ranks frames within a category.
Fog changes slowly, and single-frame flicker is noise. The pipeline smooths predictions across frames (for example, a rolling median or a small state machine that requires N consecutive frames to change category), which cuts false alarms dramatically. Temporal consistency is one of the most important engineering decisions in the project.
Evaluation reports per-class precision/recall/F1 plus a confusion matrix, with special attention to the two failure modes that matter: false fog alarms (clear day flagged as foggy) and missed dense fog. Lighting variation is called out: night frames, lens flare, and rain are documented as known failure regions, and evaluation can be split by day/night if the data supports it.
A simple dashboard shows the current category with confidence, the smoothed fog score over time, and an alert log with timestamps and frame thumbnails. Alerts are advisory: the system recommends a human look, it does not take action.
Regression to visibility distance, multi-camera aggregation, and edge deployment are natural second-phase additions.
This project’s objective — how well can this scene be seen — is distinct from every existing computer-vision Idea on the site. The AI-Powered Campus Attendance via Facial Recognition detects and matches people’s faces; the Medical Image Annotation Tool for Researchers builds labeling workflows over medical images; and the Real-Time Dashboard Builder for IoT Data Streams visualizes sensor time series. This project classifies atmospheric fog and visual visibility from camera frames — a weather-monitoring task, not a people, medical, or dashboard task — while sharing the computer-vision foundation of the first two and the time-series alerting pattern of the last. It is also the site’s first environment-domain project, pairing with the air-quality prediction Idea as the environmental-monitoring family.
| Tool type | Approach | Limitation |
|———–|———-|————|
| Professional visibility sensors | Laser/scatter instruments | Expensive; sparse deployment |
| Weather-service human observations | Trained observers | Manual, infrequent, not camera-scalable |
| Generic scene classifiers | Pre-trained models | Not trained on fog; no visibility framing |
| Commercial smart-camera systems | Integrated weather analytics | Opaque, costly, often bundled with surveillance features |
This project’s differentiators: a fully transparent, student-buildable fog classifier, hybrid learned+hand-crafted features, temporal consistency engineered for low false alarms, and an explicit advisory-only boundary.
Browse more Project Ideas · Intermediate Ideas
Build a pipeline that ingests public air-quality sensor data, forecasts PM2.5/AQI with uncertainty, and maps pollution hotspots — informational environmental...
A research tool for exploring and comparing crop trait and germplasm data from public agricultural databases — normalized schemas, provenance...
Build a crop-yield estimation tool that combines field and soil information with weather and historical agricultural data — explainable ML...
Published on September 9, 2026
A team of developers, researchers, and innovators who review and publish practical ideas for builders and creators.
Published on September 9, 2026
A team of developers, researchers, and innovators who review and publish practical ideas for builders and creators.