基于实例分割的双目特征点匹配目标识别和定位研究

Research on Binocular Feature Point Matching Target Recognition and Location Based on Instance Segmentation

针对双目视觉定位中对物体类别与距离远近判定的实际需求,提出了一种结合实例分割与特征点匹配的定位方法,准确地实现了目标的识别和定位。该方法通过Mask Region with Convolution Neural Network Feature(Mask R-CNN)对双目相机采集到的左图像特进行目标检测和分割,采用SURF算法提取分割区域的特征点并与右图特征点进行匹配得到视差,利用双目视差测距原理计算出目标相对于摄像头的位置。同时,针对相机自身标定误差造成在远距离情况下目标定位误差逐渐变大的问题,采用了最小二乘法对视差进行拟合。实验结果表明,该方法不仅能够精准实现目标识别,而且与未经拟合处理结果相比,平均误差值由0.183 m降低到0.106 m,定位精度得到了显著提高。

Based on the practical requirements for determination of object type and distance in binocular vision location,a method combining instance segmentation and feature point matching is proposed to accurately realize target recognition and location.The method firstly performs target detection and segmentation on the left image acquired by the binocular camera through Mask Region with Convolution Neural Network Feature(Mask R-CNN),then extracts the feature points of the segmentation region by SURF algorithm and matches the feature points on the right image to obtain parallax,finally,the binocular parallax ranging principle is used to calculate the position of the target relative to the camera.At the same time,to solve the problem that the camera’s own calibration error causes the target location error to become larger in a long distance,the least square method is used to fit the parallax.The experimental results show that the proposed method can not only achieve target recognition accurately,but also reduces the average error value from 0.183 m to 0.106 m as compared with the unfitting result.The location accuracy is significantly improved.