Abstract: Intelligent harvesting is often required for object detection under complex field conditions, especially in the process of the real-time monitoring of harvesting quality and accurate grading conveyor. Taking Panax Notoginseng as the research plant, this study aims to propose lightweight object detection using YOLOv5s. The complex field conditions included the large variations in the light intensity, the difficulty in separating the roots from the soil, easy entanglement of roots, variable lifting speed, vibration amplitude, and frequency. The optimal model was also obtained with the high accuracy, and low complexity of a large model suitable for the deployment of mobile terminals. Firstly, a sample dataset was collected from the Panax Notoginseng in the complex field. The influence parameters of transportation and separation were also determined for the complex root-soil system; Secondly, real-time detection was realized under complex field conditions. Slim-neck lightweight neck network was introduced into the lightweight convolution of GSConv. The original SPPF feature fusion module was retained, while the ShuffleNetv2 lightweight feature extraction network was used to improve the original backbone network, which greatly reduced the model complexity with the model accuracy; Finally, the loss function with angular penalty metric (SCYLLA-IoU, SIoU) was used to optimize the bounding box loss function, in order to enhance the detection accuracy and generalization performance of the lightweight improved model. Ablation experiments were carried out to verify three improvement strategies, namely the Slim-neck neck feature extraction network, ShuffleNetv2 backbone feature extraction network, and SIoU bounding box loss function. The experimental results showed that the improved lightweight model (PN-YOLOv5s) had 3.27×10⁶ M parameters, 5.4 G computational complexity, 6.85 MB weight size, and a detection speed of 108 frames per second. The number of parameters and weight size were approximately half of the original YOLOv5s, while the computational complexity was about one-third of the original model, and the detection speed increased by 1.2 times. Additionally, the precision of the improved model reached 93.15%, which was almost the same as the original model. The recall reached 89.46% with an improvement of 0.48 percent points, compared with the original model. The F₁ score reached 91.27% with an improvement of 0.22 percent points. The mean average precision reached 94.20%, only 0.6 percent points lower than the original. Compared with the mainstream SSD, Faster R-CNN, YOLOv4-tiny, YOLOv7-tiny, and YOLOv8s models, the improved lightweight model greatly reduced the complexity of the model, indicating better overall performance, in terms of model accuracy and real-time detection. The best performance was achieved in the actual harvesting, indicating the stronger robustness more suitable for deployment into mobile terminals. The bench tests showed that when the lifting inclination angle and vibration amplitude remained unchanged, the detection performance of the improved model decreased with the increase in lifting speeds and vibration frequencies. On the whole, the target detection of Panax Notoginseng was achieved with a precision of over 90%, an F₁ score of over 86%, and a mean average precision of over 87% under four conditions of conveying and separation operating. There was little difference in the detection speed. But the improved model can fully meet the requirements of real-time detection under actual harvesting conditions. The finding can provide technical support to the subsequent monitoring of harvesting quality and adaptive grading conveyor in the Panax Notoginseng combine harvesters.