Aiming to address the challenge of balancing detection accuracy and model lightweightness in automated passion fruit harvesting under densely occluded orchard environments, an improved YOLO 11n-based visual detection model was proposed. Partial convolution (PConv) was introduced in both the feature extraction stage and detection head to replace standard convolution blocks for lightweight optimization. A sliced 3D spatial and channel attention module ( SimAMs), integrating a feature map slicing mechanism, was designed to enhance cross-channel and spatial feature fusion. Experimental results showed that the improved model achieved a precision of 93.32% and mAP@0.5 of 93.08% , with increases of 1.28 percentage points and 0.26 percentage points over the original model. The parameter count, computation, and memory usage were reduced by 21.2% , 23.8% , and 20.0% , respectively. Detection speeds on GPU and CPU were approximately 1.36 times and 1.68 times faster than the average speed of the compared models, including EfficientNetV2 and FasterNet. Harvesting tests showed a single-fruit picking time of 13 s and a success rate of 91.7% at a depth range of 400 ~ 500 mm, 11.1 percentage points and 2.8 percentage points higher than that at near and far distances. The research result can provide effective technical support for high-precision, real-time detection and robotic harvesting of passion fruit in complex orchard environments with dense occlusion.