The accurate and robust detection of fruits in the greenhouse is a critical step of automatic robot harvesting. However, the complicated environmental conditions such as uneven illumination, leaves or branches occlusion, and overlap between fruits make it difficult to develop a robust fruit detection system and hinders the step of commercial application of harvesting robots. In this study, we propose an improved anchor-free detector called TomatoDet to deal with the above challenges. First, an attention mechanism is incorporated into the CenterNet backbone to improve the feature expression ability. Then, a circle representation is introduced to optimize the detector to make it more suitable for our specific detection task. This new representation can not only reduce the degree of freedom for shape fitting, but also simplifies the regression process from detected keypoints. The experimental results showed that the proposed TomatoDet outperformed other state-of-the-art detectors in respect of tomato detection. The F1 score and average precision of TomatoDet reaches 95.03 and 98.16%. In addition, the proposed detector performs robustly under the condition of illumination variation and occlusion, which shows great promise in tomato detection in the greenhouse.
Keywords: CenterNet; anchor-free; deep learning; harvesting robots; tomato detection.
Copyright © 2022 Liu, Hou, Liu, Liu, Zhao and Li.