Authors

Abstract

The grain depot scenario is critical for grain logistics and transportation, and it is also a key setting for the efficient operation of intelligent grain logistics platform vehicles. A large number of repetitive and specific building structures, along with low-textured walls, characterize the grain depot scene. Loopback detection is an essential module in visual SLAM, and an efficient system can eliminate accumulated errors. While traditional systems rely on manually designed features, which struggle to adapt to the unique grain depot environment, this paper proposes a deep learning-based loopback detection system for grain transfer trucks. Leveraging a custom dataset capturing both grain depot environments and loopback scenarios, the system employs convolutional neural networks for identifying building equipment and door numbers, edge extraction for robust feature matching, and image template matching for efficient loopback verification. Extensive testing on the grain depot loopback dataset demonstrates that the system significantly improves loopback detection accuracy and efficiency, paving the way for reliable autonomous navigation in grain depots.

Abstract in Chinese

粮库场景是粮食物流转运的重要场景，同时也是智能粮食物流平台车高效运行的关键环节。粮库的大量重复和特殊建筑结构以及缺乏纹理的墙体颜色是粮库场景的特点。回环检测模块是视觉定位与建图的一个重要模块，有效的回环检测能够消除累积的误差，传统的回环检测使用的特征是人工设计的特征，在粮库的特殊场景下难以发挥出良好的效果。本文提出了一种基于深度学习的粮食转运车回环检测系统，利用录入了粮库环境和回环场景的定制化数据集，使用卷积神经网络识别建筑设备和门牌号码，通过边缘提取进行稳健的特征匹配，并采用图像模板匹配进行高效的回环验证。在粮库回环数据集上进行的广泛测试表明，该系统显著提高了回环检测的准确性和效率，为在粮库中实现可靠的自动导航铺平了道路。