Overview
My system is a real-time computer vision solution built around a forward-facing camera connected to a Raspberry Pi running a custom-trained EfficientDet object detection model. The central concept is that a wrong-way driver will consistently see the backs of road signs, while a driver traveling in the correct direction will not. The model is trained to detect the backs of road signs based on their shape and reflective characteristics. All processing takes place locally on the Raspberry Pi. The camera feed is analyzed frame by frame in real time, and the system operates entirely offline without requiring an internet or cloud connection.
Training
To develop the model, I drove over 200 miles across North Texas highways and collected thousands of video frames. I primarily collected data during nighttime hours, since wrong-way driving accidents commonly occur during that time. Then, I manually annotated the dataset, labeling the backs of road signs in each frame. The goal was to achieve consistent performance not only in clear daytime conditions, but also in more challenging environments, including nighttime, rain, and glare from headlights.
Detection Algorithm
During runtime, the system does not rely on a single detection to send an alert. Instead, it uses a sliding window by checking for multiple detections within a short time interval. Once the computer vision model detects an instance of wrong-way driving, it sends a continuous audio alert to the driver that is designed to grab their attention and instruct them to pull over safely. This approach reduces the number of false positives by ensuring that a single, isolated detection does not trigger an alarm.
Latency
Because vehicles travel at highway speeds of roughly 70 miles per hour, fast processing speed is important. Using edge computing, all processing occurs locally on the device, eliminating delays and allowing detections to be made in real time. This also ensures that any alerts are delivered quickly enough to give the wrong-way driver enough time to react.