基于改进Mask-RCNN算法的车位检测研究

A research on parking slot detection based on improved Mask-RCNN algorithm

车位检测是自动泊车至关重要的环节,在复杂情况下,为同时实现自动泊车视觉系统对车位识别和车位状态分类,提出一种基于改进掩模区域卷积神经网络(Mask Region Convolutional Neural Network,Mask-RCNN)算法的C-Mask-RCNN车位检测算法。C-Mask-RCNN车位检测算法通过在Mask-RCNN算法的ResNet50特征提取网络中增加卷积块注意力模块(Convolutional Block Attention Module,CBAM),使模型更加关注车位相关的语义信息。利用C-Mask-RCNN车位检测算法中的区域卷积神经网络(Regions with Convolution Neural Network,RCNN)分支网络进行车位检测,实现Keypiont分支进行车位8个关键点的预测。实验结果表明,改进后的C-Mask-RCNN车位检测算法较Mask-RCNN算法在车位类型识别平均精确率上提升7.4%,在车位状态识别平均精确率上提升11.1%,并且车位线关键点预测的平均像素误差减少15.1 px。

Parking slot detection is a crucial part of automatic parking.In complex situations,in order to achieve automatic parking vision system for parking slot recognition and parking slot status classification.A parking slot detection algorithm of C-Mask-RCNN was proposed based on improved algorithm of Mask Regions Convolutional Neural Network(Mask-RCNN).C-Mask-RCNN adds Convolutional Block Attention Module(CBAM)to ResNet50 feature extraction network of original Mask-RCNN algorithm to make the model pay more attention to the semantic information related to parking slot.The Regions with Convolution Neural Network(RCNN)branch network of C-Mask-RCNN was used to detect parking slot,and the Keypiont branch was used to predict the 8 keypoints of parking slot.The experimental results show that the C-Mask-RCNN algorithm slot-type-detection average accuracy is improved by 7.4%,the slot-status-detection average accuracy is improved by 11.1%,and the average pixel error of keypoints is reduced by 15.1 px.