面向电商无人配送复杂场景的感知模型研究

Research on Perception Models for Complex Scenarios in E-Commerce Unmanned Delivery

摘要: 在电商高速发展与用工成本上升的背景下，无人配送具备降本增效、全天候运行与无接触服务等优势，但其规模化落地仍受限于复杂环境下的稳健感知。环境感知是保障无人配送车辆安全通行与高效决策的关键，其准确性直接决定配送车对动态场景的理解能力。基于Transformer的RT-DETR依托全局注意力与端到端检测实现了较高效率与精度，但在无人配送典型场景中的多尺度目标与频繁遮挡下，仍存在特征融合与遮挡鲁棒性不足。为此，本文提出面向电商无人配送的RT-DETR改进方案：在骨干网络关键层嵌入自适应聚焦全局上下文注意力模块，通过动态调节感受野增强多尺度表征，从而提升对小目标与遮挡体的可分辨性；并在FPN/PAN中引入指数移动平均增强的跨维度注意力机制，以更稳健地建模长程依赖并优化跨层特征融合。实验结果表明，改进模型在Udacity自动驾驶数据集上实现mAP@50提升25.6%、mAP@50-95提升13%，验证了方法在电商无人配送典型场景中的迁移性与应用价值。

Abstract: Against the backdrop of rapid e-commerce development and rising labor costs, unmanned delivery offers advantages, such as cost reduction, efficiency improvement, 24/7 operation, and contactless services. However, its large-scale deployment remains constrained by robust perception in complex environments. Environmental perception is crucial for ensuring the safe passage and efficient decision-making of unmanned delivery vehicles, with its accuracy directly determining the vehicle’s ability to understand dynamic scenes. While the Transformer-based RT-DETR achieves high efficiency and accuracy through global attention and end-to-end detection, it still suffers from insufficient feature fusion and occlusion robustness when dealing with multi-scale objects and frequent occlusions in typical unmanned delivery scenarios. To address these issues, this paper proposes an improved RT-DETR model tailored for e-commerce unmanned delivery. An adaptive global context attention module is embedded into key layers of the backbone network to enhance multi-scale representation by dynamically adjusting the receptive field, thereby improving discernibility for small objects and occluded targets. Additionally, an exponential moving average-enhanced cross-dimensional attention mechanism is introduced into the FPN/PAN to more robustly model long-range dependencies and optimize cross-layer feature fusion. The experimental results demonstrate that the improved model achieved a 25.6% increase in mAP@50 and a 13% improvement in mAP@50-95 on the Udacity autonomous driving dataset, validating the transferability and application value of the proposed method in typical e-commerce unmanned delivery scenarios.