Objective In visual simultaneous localization and mapping VSLAM when the robot or mobile device is working its pose will drift with the change of time resulting in mapping errors in some complex environments. Loop closure detection is one of the important parts of the VSLAM system which can help to converge and eliminate errors by the algorithm. Aiming at the problem of insufficient detail feature extraction in loop closure detection a loop closure detection algorithm with dual-feature fusion is proposed. Methods Firstly image features were extracted using a pretrained VGG19 convolutional neural network with partial weights and the histogram of oriented gradients HOG algorithm. The extracted two sets of features underwent principal component analysis PCA for dimensionality reduction after which they were concatenated and fully connected to obtain a fused feature vector matrix. Then the similarity scores of the resulting fused feature matrix were calculated to generate a similarity matrix. Finally the algorithm was tested on the New College and City Centre datasets. Results Experimental results on the two public datasets indicated that the proposed dual-feature fusion algorithm exhibited stronger feature recognition capabilities. Specifically it achieved a notably higher precision at a fixed recall rate of 50% demonstrating greater robustness compared to mainstream singlefeature extraction methods. Conclusion Dual features exhibit superior image representation capability compared to individual convolutional features or traditional hand-crafted geometric features thereby more effectively fulfilling the requirements of loop closure detection