Abstract: With the widespread application of unmanned vehicles, unmanned drones, and wheeled or tracked mobile intelligent agents in fields such as logistics transportation and environmental monitoring, the target search tasks in complex scenarios have placed higher demands on their environmental perception and collaborative decision-making capabilities. To address insufficient multi-scale environmental perception and low collaborative scheduling efficiency in mobile agent target search within complex scenes, this study proposes a multi-agent target search algorithm based on shunted Transformer perception and large language model scheduling optimization. For perception, the study introduces the ViT-FRCNN network and designs a multi-scale shunted Transformer module. The query branch maintains full spatial resolution, while key and value branches are downsampled in parallel at rates of 1/1, 1/2, 1/4, and 1/8. A local enhancement module based on 3×3 depthwise separable convolution is introduced to compensate for detail loss, constructing an agent target detection algorithm based on shunted Transformer perception that achieves efficient multi-scale feature aggregation. For scheduling, a large language model is introduced, and a zero-shot prompt template is designed. An iterative closed loop of assignment, execution, feedback, and reassignment is adopted to dynamically adjust tasks, and a spatial conflict detection rule based on distance threshold is introduced. A bidirectional feedback mechanism between detection and scheduling is constructed, where the Large Language Model (LLM) filters detection results according to task semantics and guides the detection module to focus on target-related regions. Consistency constraints ensure semantic alignment between detection outputs and scheduling decisions, avoiding performance degradation caused by unilateral optimization and forming a closed-loop collaborative process of detection, scheduling, feedback, and redetection. Experimental results show that on the OpenImages V7 dataset, the proposed target detection algorithm achieves an F1 score of 98.83% and a recall of 97.76%, with mAP@0.5, mAP@0.75, and mAP@0.5:0.95 of 55.47%, 36.26%, and 30.68%, respectively, all outperforming comparison algorithms. On the COCO2017 dataset, the F1 score and recall are 96.85% and 95.72%, respectively, verifying generalization. For multi-agent target search, the accuracies for chair, toothbrush, pencil, and eraser are 98.9%, 97.4%, 97.8%, and 98.3%, respectively, with an end-to-end response time of approximately 612 ms, all outperforming comparison algorithms. These results demonstrate the effectiveness and practical value of the proposed algorithm for multi-agent target search in complex scenes.

摘要： 随着无人车、无人机和轮式或履带式移动智能体在物流运输和环境监测等领域的广泛应用，复杂场景下的目标搜索任务对其环境感知与协同决策能力提出了更高要求。针对移动智能体在复杂场景下目标搜索时存在的多尺度环境感知能力不足与多智能体协同调度效率低的问题，研究提出了一种基于分流Transformer感知与大语言模型调度优化的多智能体目标搜索算法。在感知方面，引入Vision Transformer-Faster R-CNN (ViT-FRCNN)网络，设计多尺度分流Transformer模块，使查询保持全分辨率，键和值按1/1、1/2、1/4、1/8四分支并行下采样，并引入3×3深度可分离卷积的局部增强模块补偿细节损失，构建基于分流Transformer感知的智能体目标检测算法，实现多尺度特征高效聚合。在调度层方面，引入大语言模型，设计零样本提示模板，通过分配、执行、反馈、再分配的迭代闭环动态调整任务，并引入基于距离阈值的空间冲突检测规则；同时构建检测与调度双向反馈机制，(Large Language Model, LLM)依据任务语义过滤检测结果并引导检测模块聚焦目标相关区域，通过一致性约束确保检测输出与调度决策语义对齐，避免单方面优化退化，形成检测、调度、反馈、再检测的闭环协同流程。实验结果表明，目标检测算法在OpenImages V7数据集上的F1分数和召回率分别为98.83%和97.76%，mAP@0.5、mAP@0.75和mAP@0.5:0.95分别为55.47%、36.26%、30.68%，均优于对比算法；在COCO2017数据集上F1分数和召回率分别为96.85%和95.72%，验证了泛化性。多智能体目标搜索算法对椅子、牙刷、铅笔和橡皮擦四类物体的搜索准确率分别为98.9%、97.4%、97.8%、98.3%，端到端响应时间约612ms，均优于对比算法。上述结果验证了所提算法在复杂场景多智能体目标搜索中的有效性与实用价值。