Rapid Recognition and Localization Based on Deep Learning and Random Filtering

Abstract

Occlusion is a challenge for accuracy of recognition and localization for picking robot in harvesting. The algorithm developed in this paper combing deep learning and random filtering, which can be effectively eliminate the disturbing of obstacle. Firstly, apple images under occluded are as training input for neural network YOLO. Recognition result was signed by rectangular after YOLO. Secondly, for eliminate the interference of occlusion by random sampling filtered the depth values of obstacle points. Experiments shown that, this algorithm can provide real-time and accurate location information for picking. The difference between real and detection is limited in 4mm by the measure range from 600mm to 1500mm. The algorithm proposed in this paper could be flexible applied in other fruits, and it has a widely application values.