基于深度学习的盲道障碍物检测算法研究

Research on Obstacle Detection Algorithm of Blind Path based on Deep Learning

摘要:

针对盲人出行时盲道场景复杂度高,已有目标检测算法对远距离障碍物以及条形障碍物特征提取困难,造成漏检等问题提出改进。针对条形障碍物检测增加非对称卷积模块(ACB),强化网络在垂直与水平方向的特征提取；构建混合池化模块,将条形池化引入网络与金字塔池化融合为混合池化模块(MPM),增强网络对长条形与非长条形障碍物检测效果；网络末端改变特征融合方式,低级特征与高级特征相乘形式以加强复杂场景下盲道障碍物识别。实验结果表明,在盲道障碍物数据集上,改进算法对比YOLO V4在多个评价指标上都有提升；实际场景测试中对远距离障碍物以及条形障碍物检测的检测精度提升明显。

Abstract:

In view of the high complexity of blind road scene when blind people travel, the existing target detection algorithm is difficult to extract the features of long-distance obstacles and strip obstacles, resulting in missed detection and other problems. Asymmetric convolution module (ACB) was added for bar obstacle detection to strengthen feature extraction in vertical and horizontal directions. A hybrid pooling module was constructed. Strip pooling was introduced into the network and pyramidal pooling was integrated into a hybrid pooling module (MPM) to enhance the detection effect of the network on the long and non-long obstacles. At the end of the network, the fusion mode of features is changed, and the multiplication form of low-level features and advanced features is used to strengthen blind obstacle recognition in complex scenes. The experimental results show that, compared with YOLO V4, the improved algorithm has improved in multiple evaluation indexes in the blind obstacle data set. In the actual scene test, the detection accuracy of long-distance obstacles and strip obstacles is improved obviously.