Executive summary Executive summary
Redflex is an Australian-based company that develops intelligent transport solutions (ITS) for government, police, and traffic departments to make cities greener, safer, and smarter. The company has a large presence in the USA, the UK, as well as Asia and the Middle East. It offers a range of safety solutions - including speed cameras, red-light cameras, and school bus stop-arm cameras that deter dangerous driving.
Redflex wanted to increase their market presence with a new solution for traffic management. The client needed to validate their product idea and develop an intelligent transport solution with high detection accuracy.
N-iX set up a strong team of computer vision specialists and started the project with a PoC. Together with the client and its team in Australia, our experts have been working on seat-belt fastening detection and capturing distracted driving behaviors.
The N-iX team has helped the client develop a solution with Computer Vision and Deep Learning to identify an offender behind the wheel and prevent accidents on the road.
Success story in detail
Redflex decided to increase their presence in Europe and expand the portfolio of their services. Therefore, the client needed to create a next-generation solution for traffic management, taking into account the traffic rules and related policies of different countries. Redflex installs cameras on specific road parts to detect cars, their plate numbers, speed, color, and size. The main task was to develop advanced models to detect a certain anomaly based on the image received from the client's cameras. Together with the client's R&D team, N-iX worked on two main tasks:
- Person detection (capturing distracted driving behavior);
- Seat belt fastening detection/violation classification.
N-iX gathered a team of experienced computer vision specialists and started the project with guides on annotating each windshield and driver behaviors. With the help of Computer Vision and Deep Learning, we built a PoC and offered several options for neural networks with different approaches to network training, allowing the client to identify a person behind the wheel. Our team has been working on another PoC that allows capturing specific behaviors of distracted driving, such as talking or texting on the phone, eating and drinking, etc. For added security, our engineers used AWS virtual machines as work stations.
At the start of this project, we received the first data samples and determined various factors and potential difficulties that affect image quality. These include sun glare on the windshield, lack of lighting, camera location and angles of rotation, etc. We prepared the guides with descriptions of how we will annotate each windshield and capture driver behavior. As a result, we developed a system that automatically detects whether a driver has fastened the seat belt or not. The detected violations are then sent to the dispatcher, who must confirm them. We used TensorFlow to train the model and pandas software for data analysis. The solution we developed performs at approximately 88% detection accuracy.
The next PoC focused on capturing distracted driving behavior, which divert attention away from driving, such as talking or texting on the phone, eating and drinking, talking to people in the vehicle, or smoking while driving. The system includes automatic Number Plate Recognition (ANPR) and general identification of vehicles based on Artificial Intelligence (AI) algorithms. The model we developed supports a real-time stream (30 FPS) that helps generate fines in real-time based on captured data and specific violations. Our solution provides a 91% person detection accuracy both at night and during the day.
Using Computer Vision and Deep Learning techniques, we built the PoC and offered several options for neural networks with different approaches to network training. The solution has a detection accuracy of nearly 88%.
We are currently working on the next PoC to capture distracted driving behaviors, which divert attention away from driving. Despite many constraints, we managed to achieve a high detection accuracy of nearly 91% for both night and day images. Also, we trained a model that supports a real-time stream.
- Helped RedFlex validate their product idea to expand to new markets;
- Built a PoC of a system that automatically detects whether a driver has fastened the seat belt or not;
- Reached a high level of accuracy for both PoCs;
- Developed and trained a model that supports real-time streams helping create fines in real-time based on captured data.