基于MVSVM和超像素的可通行区域检测方法

Traversable region detection method based on MVSVM and superpixels

针对地面智能机器人的可通行区域检测问题,提出一种基于MVSVM和超像素的可通行区域检测方法.利用超像素块作为特征窗口,进行视觉特征的提取,解决了基于矩形块作为特征窗口一大缺陷,即同一特征窗口可能存在多个目标的问题.通过引入超像素,在像素级的尺度上,可通行区域漏检和误检率相对于矩形特征窗口方法大大降低.同时通过采用MVSVM作为分类器,解决了传统的单平面SVM分类器在大规模分类问题中存在的需要较高内存和计算代价以及无法解决例如异或问题等复杂分类问题.在野外实际环境下的实验表明:本方法在可通行区域检测准确率上较以往方法有大幅度提高,能够较好地完成复杂场景下的可通行区域检测.

In order to detect the traversable region for ground intelligent robot,a traversable region detection method based on multi-weight-vector projection support vector machine(MVSVM)and superpixels was proposed.Superpixels were used as patches to extract the appearance-based features.The superpixel patches could overcome the disadvantage of the fixed-sized patches which contained more than one important object in one patch.By introducing superpixels,the FP and FN in traversable region detection could be greatly reduced in the pixel-level compared with the fixed-sized patches based methods.Meanwhile,using MVSVM as the classifier solved the problems such as high memory demand,high computational cost and XOR problem existed in the traditional single plane SVM classifier in large-scale classification.The experiments in actual complex wild environment demonstrate that the proposed method can obtain much higher accuracy than the methods used before on the traversable region detection.