Self-localization is the premise of navigation and other intelligent behavior for mobile robot. This paper proposes a kind of binocular vision odometer based on the bundle adjustment method. The feature point set is matched based on the point pattern matching method, and the corresponding 3D coordinate is calculated through 3D reconstruction. Then, the relative motion parameter is optimized through segmenting the image sequence into several subsequences according to overlapped regions. The performance of the binocular vision odometer does not depend on the count of rotating shaft, so that the effect of wheel radius, idle running and wheel slipping on the measurement precision can be avoided, and the relative positioning accuracy is higher.