Object detection is a prerequisite task for behavior recognition of passengers at smart ports. To solve the issues of occlusion caused by crowded queues and turnstile equipment at fast lanes, which lead to a decrease in the detection accuracy of existing models, and to meet the requirements of edge devices, an improved YOLOv12n-based occluded object detection algorithm, MSS-YOLO, is proposed. First, an Adaptive Multi-scale Feature Extraction Module (AMFE) is designed in the backbone network to replace the downsampling module. It expands the model"s receptive field and adaptively focuses on key human features. This enhances multi-scale feature extraction and reduces the loss of information and occlusion interference. Next, an improved feature enhancement module, A2C2f_SEAM, is added to the neck network. It learns the relationship between occluded and non-occluded areas and adjusts the focus dynamically. Finally, CIoU is used to replace Alpha-Shape IoU. Power transformation and geometric constraints (shape and location penalty terms) are added to solve the limitations of traditional IoU loss in occlusion scenarios and improve overall network performance. The model is experimentally verified on the public dataset (CrowdHuman) and the self-collected port crossing behavior dataset (CBCD). The results show that the MSS-YOLO model outperforms other classical models in several key performance metrics. Compared to the baseline model, precision increases by 1.5%, recall by 7.1%, mAP@0.5 by 5.4%, and mAP@0.5:0.95 by 4.5%. The model also reduces the number of parameters by 0.15×10^6. The MSS-YOLO algorithm achieves high accuracy in occluded object detection in port customs scenarios while maintaining lightweight characteristics, ensuring it meets the deployment requirements of port edge devices.