基于改进SSD-MobileNet算法的AGV动态目标检测方法

为了提升自动导引运输车(automated guided vehicle，AGV)动态障碍物视觉检测的精度和帧率，提出了一种基于单镜头多盒检测器（single shot multibox detector，SSD）的改进算法。将轻量级MobileNet网络引入到SSD网络结构中，然后利用K-means算法对训练数据集中真实框的AR值进行聚类并更新，最后利用Jeston Nano嵌入式平台搭建了AGV实验系统，引入TensorRT加速引擎，分别对改进前后的SSD-MobileNet模型进行加速优化，并对比分析。结果表明：改进的SSD-MobileNet模型在AGV上使用TensorRT加速引擎的mAP值为79.1%，相比优化前提升了10.8%，对精度影响很小，而帧率达到了25 f/s，较原SSD模型提升了近4倍，且改进后模型规模也比优化前缩小了37%。采用改进算法能够使AGV在运输过程中完成动态障碍物检测任务，可代替人工实现货物高效运输，并节省运输成本，为智能化运输提供了一种新的思路。

AGV dynamic targets detection method based on improved SSD-MobileNet algorithm

In order to improve the accuracy and frame rate of dynamic obstacle visual detection for automated guided vehicle (AGV), an improved algorithm based on the single shot multibox detector (SSD) was proposed. The lightweight MobileNet network was introduced into the SSD network structure, and then the K-means algorithm was used to cluster and update the Aspect Ratio (AR) values of the real boxes in the training dataset. Finally, an experimental AGV system was built on the embedded Jeston Nano platform and the TensorRT acceleration engine was introduced to optimize the SSD-MobileNet algorithm before and after improvement, and then comparative analysis was made. Experimental results show that the improved SSD-MobileNet algorithm has a mean Average Precision (mAP) value of 79DK(].DK)]1% on the AGV using the TensorRT acceleration engine, an increase of 10DK(].DK)]8% compared to that before optimization, with little impact on accuracy. The FPS frame rate reaches 25 frames per second, which is 4 times higher than that with the original SSD algorithm, and the model size after improvement is also 37% smaller than that before optimization. The improved algorithm can enable the AGV to complete dynamic obstacle detection tasks during transportation, which can replace the manual transport of objects efficiently and save the transportation cost, and provides a new idea for intelligent transportation.