Drivable Area Obstacle Detection for Unknown Road Objects

Instead of training the model to detect what objects are present (positive classification), the patent inverts the approach by training the model to detect what is NOT on the road surface (negative classification). The drivable area model learns to identify pixels that belong to the road surface, and anything not classified as road surface is treated as a potential obstacle. This inversion allows the model to detect any object regardless of its class, solving the problem of generic obstacle detection.

The patent creates a universal detection framework where a single drivable area model serves multiple purposes: it can detect vehicles, pedestrians, animals, debris, or any other object on the road surface without requiring separate detection mechanisms for each object type. The model's output is used by both an object detection module and a generic obstacle detection module, making the system multi-functional and adaptable to various scenarios.

The system eliminates the need for manual annotation by using an unsupervised learning approach. The drivable area model automatically learns to distinguish road surface from non-road surface elements through self-supervised training, where the training data does not require manual labeling of object classes. The model serves itself by automatically adapting to detect any object type without human intervention in data annotation, significantly reducing time and cost.

Instead of collecting and annotating real-world data for every new object class, the patent uses synthetic training images generated by placing obstacle images onto road surface images. This copying approach creates artificial training data that teaches the model to detect obstacles without requiring actual collection and annotation of diverse real-world obstacle data, reducing the time and cost burden.