As the foundation to realize the autonomous movement of mobile robots, simultaneous localization and mapping(SLAM), which is mainly used to solve the problem of mobile robots mapping and navigation in unknown environment, has been paid more attention in recent years. Loop closure detection, one of the key steps of visual SLAM, plays an important role to make a globally consistent map and reduce accumulated error of robot pose. Current methods for loop closure detection are vulnerable to environmental influence because they always adopt traditional features such as SIFT and SURF. To improve the accuracy and robustness of loop closure detection, a method based on unsupervised Stacked Convolutional Autoencoders(CAEs) model is proposed. The trained CAEs convolution neural network is used to learn from input images, while the output features are used for loop closure detection. The results of experiment show that the proposed method, compared with traditional BoW-based methods and other methods based on deep learning model, can effectively reduce the dimension of image features and improve the effect of feature description. Thus, it can attain better accuracy and robustness in loop closure detection of robot SLAM.