Simple Stereo Matching Algorithm for Localising Keypoints in a Restricted Search Space

Modern stereo matching algorithms generally rely on matching image features such as colour and texture to find corresponding matches between images. They can provide very good results but be very computationally intensive. However, in cases where the objects to be localised lie approximately on a known plane, a much simpler algorithm can be applied. One such case is localising kiwifruit in modern orchards, as the plants are trained to grow in a planar structure known as a pergola. The proposed algorithm uses tight distance limits (based on orchard geometry) to reduce the search space for matching fruit to a small window. For the majority of kiwifruit, the search window is small enough to contain only one fruit in the adjacent image, giving only a single solution. In cases where there are tightly grouped fruit or false positive fruit detections adjacent to true positive fruit detections, there can be multiple potential matching solutions. To solve these, each potential solution is evaluated based on how closely it conforms to the mean object distance from camera and a solution is selected. On real world test data containing 121 image pairs, the algorithm has a 99.2 % true positive rate. Computation time was 1.97 ms per image pair.