A survey of computer vision methods for locating fruit on trees

A review of previous studies to automate the location of fruit on trees using computer vision methods was performed. The main features of these approaches are described, paying special attention to the sensors and accessories utilized for capturing tree images, the image processing strategy used to detect the fruit, and the results obtained in terms of the correct/false detection rates and the ability to detect fruit independent of its maturity stage. The majority of these works use CCD cameras to capture the images and use local or shape-based analysis to detect the fruit. Systems using local analysis, like intensity or color pixel classification, allow for rapid detection and were able to detect fruit at specific maturity stages, i.e., fruit with a color different from the background. However, systems based on shape analysis were more independent of hue changes, were not limited to detecting fruit with a color different from the color of the background; however, their algorithms were more time consuming. The best results obtained indicate that more than 85% of visible fruits are usually detectable, although with CCD sensors there were a number of false detections that in most cases were above >5%. The approaches using range images and shape analysis were capable of detecting fruit of any color, did not generate false alarms, and gave precise information about the fruit three-dimensional position. In spite of these promising results, the problem of total fruit occlusion limits the amount of fruit that can be harvested, ranging from 40 to 100% of total fruit, depending on fruiting and viewing conditions. This fact seriously affects the feasibility of future harvesting robots relying on images that do not contain a high percentage of visible fruit. Therefore, new techniques to reduce total occlusion should be studied in order to make the process feasible.