Automatic fruit recognition: a survey and new results using Range/Attenuation images

2 SUMMARY A review of different vision systems to recognize fruits for automated harvesting is presented. This survey of recent works in this field should be useful to researchers in this interesting area. Current research proves the feasibility of practical implementations of these computer vision systems for the analysis of agricultural scenes to locate natural objects under difficult conditions. Some basic considerations about the distributions and characteristics of the fruits in natural orange crops are discussed. The research reported here explores the practical advantages of using a laser-rage finder sensor as the main component of a 3-dimensional scanner. This sensor supplies two sources of information, the range to the sensed surface and the attenuation occurred in the round-trip travel. A model of the attenuation process is presented and used to restore images and to derive additional information: reflectance, apparent reflectance, range precision and the range standard deviation. The apparent reflectance image and the range image are used to recognize the fruit by color and shape analysis algorithms. The information obtained with both the methods is merged to find the final fruit position. The 3-dimensional information with its precision, the size and the average reflectance of the image is the final information obtained for every fruit. This information allows a selective harvesting to improve the quality of the final product for the fresh fruit market. Some experimental results are presented showing that approximately 74% of the green fruits are detected and this correct location rate is improved as the amount of mature fruits in the scene increases, reaching a 100% of correct detection over the visible fruits. No false detections were found in the test images used. Future work could be directed to extract more shape information from the range image to improve the detection results. The integration of the recognition methods with the AGRIBOT harvesting system will be reported in future publications. 3 ABSTRACT An automatic fruit recognition system and a review of previous fruit detection work are reported. The methodology presented is able to recognize spherical fruits in natural conditions facing difficult situations: shadows, bright areas, occlusions and overlapping fruits. The sensor used is a laser range-finder giving range/attenuation data of the sensed surface. The recognition system uses a laser range-finder model and a dual color/shape analysis algorithm to locate the fruit. The 3-dimensional position of the fruit, radius and the reflectance are obtained after the recognition stages. Results …