Video Analysis with Computer Vision – Ultimate Format for Traffic Sign Inventories

Automation is interesting in almost every repetitive or simple task; traffic sign mapping and inventories are an excellent use case to be automated. Digital maps are getting more and more important in the days of connected vehicles and navigation, and smart phones have become a norm. Combining these ideas, with Vaisala Road AI technology turns mobile video into an ultimate format for vehicle based computer vision in various use cases.

Vaisala’s computer vision team has developed one the most advanced traffic sign mapping systems to support asset management and traffic system safety monitoring. The RoadAI system is for those who need to monitor signs and traffic arrangements with a simple and mobile video recording system.

Filter Out False Positives

Typically, false positives are causing challenges particularly in computer vision and machine learning applications. When our technical team first illustrated how false positives are filtered out from the computer vision results before visualizing the output, I thought “This is great!”. The solution is simple, effective and reliable as most good solutions usually are.

Development of the system was initiated by the Finnish Transportation Agency in 2014. The main driver was a need to manage traffic sign inventory with images and have them mapped for navigation purposes cost efficiently and reliably. Innovative solution is to combine smartphone application and computer vision for extracting and mapping signs from a video recorded from a regular maintenance vehicle. RoadAI is now operational and is capable to map signs and signaling on roads and rail. The unique system is simple to use and interface supports, not only sings on map and individual sign images, but creation of proprietary street view services for internal use. RoadAI provides the easiest way to create your road view, and map signs and condition of roads in a fraction of the time when compared to existing methods.

After this, it is hard to claim that a system based on still images could ever be as good or reliable as a video based system.

The system is currently serving multiple city customers and is piloted by multiple DOTs. The roadmap for future includes increased use of artificial intelligence and computer vision for analyzing condition of road markings, detection of environmental phenomena and mapping of other infrastructure assets.

Map Traffic Signs Quickly and Reliably

To help you get an impression of video's superiority in confidence of the detection, let´s concentrate to the image below. It shows how a 80 km/h traffic sign has been detected altogether 21 times by a single video at a normal highway speed. It is natural that classification is incorrect in some frames; computer vision is not perfect, just like no man is. The light condition, motion blur and blowing snow are typical challenges for detection and classification. The case sign here is detected for the first time from 95 meters and for the last time from a couple of meter distance. There are altogether 3 incorrect classifications and 18 correct ones. Because the video based, 3D point cloud system can localize all detections to the same position, the 3 incorrect detections are filtered out and confidence of sign classification is superior. The confidence for the final detection and classification is the sum of the 18 individual detections; results are reliable and basically never wrong.