In the domain of autonomous driving, lane instance classification is paramount for decision-making processes within vehicles. However, conventional methods frequently necessitate the pre-determination of the number of lane lines, a practice that is ill-suited to address the demands of intricate and ever-changing road scenarios. This paper puts forward a novel lane instance classification approach, underpinned by the concept of neighbour inner product. The proposed method first extracts lane lines from the semantic segmentation mask, then represents the lane lines with coordinate points, and finally classifies the lane line instances by calculating the inner product of vectors between adjacent coordinate points. Experimental results demonstrate that, in comparison with alternative methods, this approach exhibits higher precision, recall and F1 score, and is better able to adapt to complex road environments, thus providing a novel solution for lane instance classification in autonomous driving.