Simultaneous localization and mapping (SLAM) solves the problem of localization and map construction in unfamiliar environments. Loop closure detection is a key step in SLAM and is vital to building a consistent map. Current loop closure detection is usually based on traditional features and easily affected by the external environment. In order to improve the accuracy and robustness of loop closure detection, a loop closure detection method based on convolutional neural network and sequence matching is proposed. The trained neural network model is used to extract image feature. Vector of locally aggregated descriptors (VLAD) is used to encode features as image description. Sequence matching is used to construct similarity matrix for loop closure detection. The experimental results show that compared with the method based on traditional features or deep learning model, the algorithm can achieve better accuracy and robustness.