Traversable Region Detection Based On Machine Learnign And Information Fusion

Traversable region detection is the key technique in driver assistance systems and autonomous navigation systems. Vision-based traversable region detection has been attracting attention in recent years because of its rich information and low cost. Current research is mainly based on pattern recognition and geometry. And usual features include color, texture, shape model and vanishing point etc. Images are easily influenced by scene variation and complex illumination, which means the same scene would be in different patterns with different time and illumination. This makes extreme variation to features, and brings more difficulty to stable detection. Though a lot of work has been done to overcome these diffculities, problems like poor adaptability, illumination robustness and stability are still unsolved.This paper focuses on robust traversable region detection based on environment perception system of unmanned platform and presents a traversable region detection system based on imbalanced learning mechanism. To guarantee the stability and adaptability, this system takes several noval strategies:1. Weak assumption. We only assume a small region in front of the platform is traversable. This weak assumption makes this method adaptive to more scenes.2. Feature extraction and recognition with multi-scale superpixel segmentation. A multi-scale segmentation stratgy is presented to extract complete information and assure the stability of training and recognition.3. Imbalance learning. We set weight based on a function, and build the forgetting mechanism to replace outdated samples with new ones, which increases the adaptibility to changing environment.This paper also extends the above learning stratgy by fusion with information from other sensors, which improves performance of traversable region detection system.