Autonomous Navigation of Orchard Harvesting Robot via Light-weight Salient Object Detection

• Proposing a multi-level dense feature fusion strategy to achieve features fusion. • A diverse feature aggregation module is introduced to obtain different scale receptive fields. • A light-weight SOD architecture is constructed to produce an accurate navigation line. • The average navigation line deviation is only 4.8 cm in field experiment. Autonomous navigation technology of orchard harvesting robots is a key technology which can significantly improve harvest efficiency and reduce labor. Limited by complex environment with closed orchards and severe obstruction, traditional GNSS navigation methods are susceptible to interference, resulting in signal loss. LiDAR navigation as another mainstream navigation method, is difficult to achieve real-time navigation on embedded devices because of the substantial volume of 3D point cloud information. To overcome these problems, a light-weight salient object detection (SOD) architecture driven by deep-learning techniques is developed to obtain orchard road information for autonomous navigation of orchard harvesting robot. Specifically, a diverse feature aggregation (DFA) module in the architecture is designed to obtain different scale receptive fields. To achieve interaction of multi-level features, a multi-level dense feature fusion (MLDFF) is introduced to achieve features fusion. Finally, the salient road mask obtained by the architecture is utilized to generate navigation line by using line scanning method and least squares method. Comparative tests demonstrated that our approach surpassed two conventional methods and five state-of-the-art deep-learning algorithms in detection performance. The size of our model is only 16.0 MB and the detection efficiency is 115.25 Fps, which can meet real-time detection and requirements for embedded deployment. Moreover, in terms of navigation accuracy, the centerline means the pixel-level deviations and mean heading angle errors of the navigation line reach 5.95 pixels and 3.05°, respectively. Field experiment showed that the maximum and average navigation line deviation are 10.5 cm and 4.8 cm, respectively.