Keypoint Detection and 3D Localization Method for Ridge-Cultivated Strawberry Harvesting Robots

With the development of intelligent modern agriculture, strawberry harvesting robots play an increasingly important role in precision agriculture. However, existing vision systems face multiple challenges in complex farmland environments, including fruit occlusion, difficulties in recognizing fruits at varying ripeness levels, and limited real-time processing capabilities. This study proposes a keypoint detection and 3D localization method for strawberry fruits utilizing a depth camera to address these challenges. By introducing a Haar Wavelet Downsampling (HWD) module and Gold-YOLO neck, the proposed method achieves significant improvements in feature extraction and detection performance. The integration of the HWD module effectively reduces image noise, enhances feature extraction accuracy, and strengthens the method’s ability to recognize fruit stems. Additionally, incorporating the Gold-YOLO neck structure enhances multi-scale feature fusion, improving detection accuracy and enabling the method to adapt to complex environments. To further accelerate inference speed and enable deployment in an embedded system, Layer-adaptive sparsity for Magnitude-based Pruning (LAMP) technology is employed, significantly reducing redundant parameters and thereby enhancing the lightweight performance of the model. Experimental results demonstrate that the proposed method can accurately identify strawberries at different ripeness stages and exhibits strong robustness under various lighting conditions and complex scenarios, achieving an average precision of 97.3% while reducing model parameters to 38.2% of the original model, significantly improving the efficiency of strawberry fruit localization. This method provides robust technical support for the practical application and widespread adoption of agricultural robots.