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

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18.4 MB的情况下达到82.9%的平均精度,改进的YOLOv5m在增大模型为45.5 MB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。

During the process of building stairs by using mobile robot visual simultaneous localization and mapping(SLAM),it is necessary to detect and recognize the features of stairs.Traditional stair detection technologies,such as edge detection and line extraction,often have the characteristic of ideal visual angle and simple background,but they can not extract the stair features under the railing occlusion and complex backgrounds.In order to solve the above problems,a stair target detection method based on improved YOLOv5 used for mobile robots is proposed.The FenceMask data enhancement strategy at the input end is introduced to increase the number of training samples for occluded stairs.The channel attention module(CAM)and spatial attention module(SAM)are connected in parallel to form the convolution block attention module(CBAM),enhancing the ability to extract stair features in complex backgrounds.At the prediction end,the non-maximum(NMS)and weighted boxes fusion(WBF)are combined,and the high confidence and close position bounding boxes filtered by the NMS are fused into new bounding boxes,improving the detection speed of single segment and multi-step stairs in the Faster-RCNN and single short multi-box detector(SSD)detection algorithms while meeting accuracy requirements.Simulation results show that the improved YOLOv5 reaches an average accuracy 82.9%with a model size of 18.4 MB,and improves the average accuracy of 86.5%with a model size of 45.5 MB.The improved YOLOv5 can effectively identify the conditions of railing occlusion,complex backgrounds and single segment long stairs.