Traffic Flow Analysis with Computer Vision

The problem

In Hawaii, traffic-efficiency data is gathered by people counting vehicles by hand, which is slow and labor-intensive. I talked with the head of traffic at Hawaii’s Department of Transportation about what they needed, and set out to automate the counting with computer vision.

A second problem sits underneath: many existing traffic-efficiency models depend on a time hyperparameter, which limits how flexibly they apply to different roads and periods.

Approach

Capturing the video

I filmed traffic from a tripod, planning the camera position against the path the cars take through the lot.

Planning the camera position and car path — Video curation: the path of the cars on an aerial map, the tripod setup and its view, and the resulting car-flow video.

Detection and tracking

I ran a YOLOv8 convolutional neural network on mosaic imaging, which stitches several high-resolution frames into one composite, so the network can detect and track vehicles across the whole scene at once.

How the detector finds a car — Detecting a car in a video: the frame is split into a grid, candidate boxes are scored by intersection over union against anchor boxes, and the surviving boxes become detections.

An efficiency equation without the time hyperparameter

With my brother, I wrote a new traffic-efficiency equation that removes the time hyperparameter. We extended it into a two-equation system built for simulation, which minimizes the interquartile range of flow while maximizing efficiency. It supports simulation-based planning of traffic flow. We are working toward a provisional patent on the equation, so I am not publishing it here.

Results

I field-tested the system for six months, collecting and analyzing traffic patterns.
After deployment, Hawaii’s head of traffic at the Department of Transportation reported that it cut traffic-data labor by more than 80%.
It improved the efficiency of Waipahu High School’s traffic methodology by 17%.
The project was a finalist at the International Science and Engineering Fair (about 1,600 spots from roughly 7 million students worldwide) and won two special awards there.

My contribution

I built the computer-vision pipeline, did the filming and six months of field testing, and ran the analysis. My brother and I developed the efficiency equation together.