LDMamTrack: A lightweight object detection model based on Mamba architecture for e-commerce logistics images

In the object detection of e-commerce logistics images, traditional models generally face a core contradiction: ”pursuing detection accuracy will sacrifice inference speed, while ensuring real-time performance will compromise accuracy”. To address this issue, this paper proposes a lightweight object detection model LDMamTrack based on the Mamba architecture, which achieves the coordinated optimization of speed and accuracy through three core modules: The LDMambaBlock adopts Hilbert Scan linearization and Linear Deformable Convolution (LDConv), enabling long-range feature modeling with linear complexity while adapting to deformed objects; The Simple Stem replaces the traditional large-kernel convolution with stacked small-kernel convolutions, realizing lightweight extraction of initial features of small objects; The Vision Clue Merge (VCM) module reduces redundant computation and optimizes feature transmission efficiency through dimension splitting and normalization removal design. Experimental results show that LDMamTrack achieves an mAP50 of 78.3% and an mAP50-95 of 52.0% on the LOCO dataset, and an mAP50 of 88.9% and an mAP50-95 of 56.0% on the SKU-110K dataset. Meanwhile, the model’s inference speed reaches 45 FPS. LDMamTrack can meet the needs of real-time detection in e-commerce logistics sorting lines and accurate warehouse inventory, providing technical support for the intelligent upgrading of logistics automation systems.