Stereo vision with texture learning for fault-tolerant automatic baling

This paper presents advances in using stereovision for automating baling. A robust classification scheme is demonstrated for learning and classifying based on texture and shape. Using a state-of-the-art texton approach a fast classifier is obtained that can handle non-linearities in the data. The addition of shape information makes the method robust to large variations and greatly reduces false alarms by applying tight geometrical constraints.

The classifier is tested on data from a stereovision guidance system on a tractor. The system is able to classify cut plant material (called swath) by learning it's appearance. A 3D classifier is used to train and supervise the texture classifier.