To address the issues of low detection accuracy and speed in apple tree trunk recognition, this paper proposes a precise apple tree trunk recognition method based on an improved YOLO v8 model. First, a depth-sensing camera is used to capture images of apple tree trunks, and YOLO v8 is adopted as the baseline model. The convolutional layers are replaced with re-parameterized convolution structures to enhance the model′s feature learning capability. Second, the feature fusion unit is optimized by introducing a dynamic head detection mechanism, which improves both detection speed and accuracy. Finally, field experiments were conducted using traditional YOLO v8, Fast R-CNN, and other models as baselines, with average recognition accuracy and frame rate as evaluation metrics. The results show that the improved model is capable of accurately recognizing apple tree trunks, achieving an average recognition accuracy of 95.07% and a detection speed of 112.53 f/s, and the model parameters amount to 4.512×107. Compared with the traditional YOLO v8 model, the average recognition accuracy increased by4.98 percentage points, and detection speed increased by 3.24 f/s. Compared with mainstream object detection models such as Fast R-CNN, YOLO v7, YOLO v5, and YOLO v3, the improved model outperformed them in average recognition accuracy by 15.26, 6.33, 9.59, and 13.41 percentage points, respectively, and in detection speed by 96.81, 75.27, 2.23, and 57.10 f/s, respectively. Additionally, the model′s parameter count was reduced by 9.198× 107, 1.93×106, and 1.641×107 compared to Fast R-CNN, YOLO v5, and YOLO v3, respectively. This research provides technical and methodological support for autonomous navigation and intelligent operations in apple orchards.