Security screening is a crucial method for preventing violent terrorist attacks and potential contraband threats at transportation hubs. Terahertz (THz) imaging technology, as an emerging detection technique, offers a new solution for public security screening due to its nondestructive penetration, low photon energy, and molecular fingerprinting capabilities. In this article, knives of different shapes made of metal, ceramic, and plastic are used as concealed objects, and materials such as multilayer cardboard, polyethylene bubble wrap, foam boxes, and textiles served as covering layers. A dataset of THz images containing concealed objects with mixed occlusion layers is constructed. After analyzing the THz dataset, we propose a you only look once (YOLO) detection model based on boundary information enhancement (BIE-YOLO). This model introduces omni-dimensional dynamic convolution into the backbone network to enhance the feature extraction capability. A BIE module is added to capture detailed boundary information from the shallow layers of the backbone network, mitigating missed or false detections caused by blurred boundaries in THz images. Moreover, to improve detection accuracy for objects with similar contours, a BiFormer is used in the network neck to model global information and capture long-range dependencies. Experimental results demonstrate that the BIE-YOLO model achieves F1 score, mAP50:95, and mAP50 of 79.2%, 79.9%, and 93.8%, respectively, representing improvements of 3.0%, 2.2%, and 1.5% compared to the baseline model YOLOv8. This model effectively detects concealed objects under multilayer occlusions, providing new insights and directions for applying THz technology in security screening.