一种基于并行FCN的室内视觉定位算法

Indoor vision positioning method based on parallel-FCN network

对传统的基于视觉物体识别的室内定位算法中物体识别不准确导致定位精度低的问题,提出了一种改进的并行全卷积神经网络(Parallel-Fully Convolutional Neural Network,P-FCN)物体识别算法。该改进算法通过将传统FCN网络并行化,实现在样本不平衡的训练情况下,网络能够对图像中的物体进行像素级的准确分类,从而提高物体轮廓的识别精度。与已有的区域卷积网络-模板匹配联合检测(Parallel R-CNN Detection and Template Matching Refinement,PDTR)算法和传统FCN算法相比,并通过相同验证数据集进行验证,实验结果表明,所提改进算法提高了图像中物体识别的准确率、轮廓识别精度和粗定位准确性,进而提高了最终的定位成功率和定位准确率。

In the traditional indoor positioning algorithm based on visual object recognition,the inaccurate object recognition is an important factor restricting the positioning accuracy.An improved parallel-fully convolutional neural network(P-FCN)object recognition algorithm is proposed to solve this problem.The improved algorithm parallelizes the traditional FCN network to enable the network accurately classifying the objects in the image at the pixel level in the case of unbalanced training samples,so as to improve the recognition accuracy of object contour.Compared with the existing parallel R-CNN detection and template matching refinement(PDTR)algorithm and the traditional FCN algorithm,the improved algorithm improves the accuracy of object recognition,contour recognition and coarse positioning in the image,and then improves the final positioning success rate and positioning accuracy.