Pineapple harvesting in natural orchard environments faces challenges such as high occlusion rates caused by foliage and the need for complex spatial planning to guide robotic arm movement in cluttered terrains. This study proposes an innovative visual detection model, Yolov10n-SSE, which integrates split convolution (SPConv), squeeze-and-excitation (SE) attention, and efficient multi-scale attention (EMA) modules. These improvements enhance detection accuracy while reducing computational complexity. The proposed model achieves notable performance gains in precision (93.8%), recall (84.9%), and mAP (91.8%). Additionally, a dimensionality-reduction strategy transforms 3D path planning into a more efficient 2D image-space task using point clouds from a depth camera. Combining the artificial potential field (APF) method with an improved RRT* algorithm mitigates randomness, ensures obstacle avoidance, and reduces computation time. Experimental validation demonstrates the superior stability of this approach and its generation of collision-free paths, while robotic arm simulation in ROS confirms real-world feasibility. This integrated approach to detection and path planning provides a scalable technical solution for automated pineapple harvesting, addressing key bottlenecks in agricultural robotics and fostering advancements in fruit-picking automation.