A Lightweight and Efficient Detection Transformer for Highway Abandoned Objects

Abandoned objects on highways seriously threaten traffic safety, and their prompt identification and removal are crucial. Existing methods struggle to balance computational cost and detection accuracy due to the significant scale differences of abandoned objects on highways. To address these problems, we propose a Lightweight and Efficient Detection Transformer for highway abandoned objects (LE-DETR). This study first designs a real-time feature extraction module that effectively captures essential information and accelerates information flow. Building on this module, we construct a lightweight backbone network for feature extraction, enhancing parameter utilization. A Triple Fusion (TFusion) module is proposed, integrating high-level semantic information with low-level spatial information to increase detailed information. A Cross-Layer Multi-Scale Interaction (CMI) module is designed, utilizing large-kernel depth-wise convolutions of various sizes to extract features from different receptive fields, enhancing the multi-scale representation of abandoned objects. The LE-DETR model is trained and evaluated using a constructed Highway Abandoned Object Dataset (HAOD). The experimental results indicate that compared to the suboptimal RT-DETR-R18, LE-DETR improves accuracy by 6.5%, reduces the number of parameters by 27.1%, and decreases floating-point operations (FLOPs) by 21.1%. These improvements demonstrate the great potential of LE-DETR for detecting abandoned objects on highways.