Aiming at the problem of information loss in image feature representation of loop closure detection, a feature extraction algorithm based on Vision Transformer (ViT) with convolutional neural network for multi-model fusion was proposed. Firstly, feature extraction was carried out on the input image, and then the high-dimensional image feature vector was reduced by kernel principal component analysis (KPCA) to construct a new image feature representation. At the same time, a new range-matching algorithm was proposed, which limited and selected the range for feature matching through the corresponding range framework. The experimental results show that the proposed algorithm compared with other algorithms has higher accuracy and matching rate, and achieves better robustness and real-time requirements, which proves the effectiveness of the proposed algorithm in loop closure detection.