基于改进YOLOv5的室内楼梯检测方法研究

Study on indoor staircase detection method based on improved YOLOv5

摘要:

移动机器人视觉SLAM的楼梯建图过程需要对楼梯特征进行检测识别，传统的边缘检测、直线提取等楼梯检测技术往往视角较为理想、背景较为简单，无法实现栏杆遮挡、复杂背景下的楼梯特征提取。为了解决以上问题， 提出了一种可用于移动机器人的改进YOLOv5的楼梯目标检测方法，在输入端引入FenceMask数据增强策略，增加对遮挡楼梯的训练样本数量。通道注意力模块CAM与空间注意力模块SAM采用并行连接的方式组成注意力模块CBAM，加强在复杂环境下对楼梯的特征提取能力。在预测端将NMS与WBF结合，将NMS筛选之后置信度较高且位置相邻的边框进行融合为新的边框，在满足精度要求的情况下改善了Faster-RCNN与SSD检测算法存在的单段多阶楼梯检测速度问题。仿真表明改进的YOLOv5s可以在模型大小18.4MB的情况下达到82.9%的平均精度，改进的YOLOv5m在增大模型为45.5MB的情况下平均精度提高为86.5%，均可有效识别栏杆遮挡、复杂背景以及单段长阶梯。

Abstract:

Detection and recognition of the stair feature are needed in the mobile robot visual SLAM mapping. Traditional stair detection technologies, such as edge detection and line extraction, often used ideal visual angle and simple background. They cannot finish the stair feature extraction under the condition of the railing occlusion and complex background. In order to solve the above problems, a stair target improved detection method based on YOLOv5 used for the mobile robot is proposed in the paper. The FenceMask data enhancement strategy is introduced to increase the number of training samples for occluded stairs at the input end. The feature extraction ability of the stairs in complex environment is strengthened by introducing CBAM attention mechanism. The problem of "multiple inspection" of the single multi-step stairs occurring in Faster-RCNN and SSD are solved and improved by combining NMS and WBF at the prediction end. The simulation results show that an average accuracy 82.9% can be obtained under the condition of the model size 18.4MB in the improved YOLOv5. And the average accuracy of the improved yolov5m can be enhanced to 86.5% when the model size is 45.5MB. The improved YOLOv5 method in the paper can effectively identify railing occlusion, complex background and single long steps.