Unstructured road environment is variable and unstable, but vegetation on both sides of road is more remarkable, which can be used to confine impassable area. In complex outdoor environment, vegetation area detection is vulnerable to multiple disturbance factors such as weather, shadow, road condition, and so on, resulting in detection error. Therefore a method of vegetation detection based on Gaussian kernel SVM(support vector machine) is proposed. Firstly, the sample feature of multidimensional color space is analyzed and learned through the sparse representation based on superpixel. Then, classification criteria are created for effectively absorbing vegetation information. Also, grid probability filtering are used to optimize testing results and improve the detection accuracy. Experiments show that the approach excellently solves the vegetation detection problem in unstructured road environment, which is of strong anti-interference ability facing the changing lighting and road condition, and has superior real-time performance and reliability. In practical applications, impassable regions on road are effectively restricted, ensuring the security area of the intelligent mobile robot in complicated road environment.