Abstract: The detection of loop closure is a very important part of visual slam. Successful detection of loop closure can reduce the accumulated mileage drift generated by positioning algorithms. In view of the superior performance of deep convolutional neural networks in classification, the network of VGG16-Places 365 is used, which is widely used in image classification to the area of loop closure detection for the first time. The registration data are input into a trained convolutional neural network, and the output of each hidden layer corresponds to the image feature representation. Then, experiments are implemented to get an intermediate layer with higher matching accuracy, which is used to complete scene feature extraction, and then the loop closure region is obtained by calculating the similarity of the scene feature; finally, experimental tests are performed on loop closure detection dataset. Test results show that the accuracy rate of the VGG16-Places 365 convolutional neural network model is about 3% higher than the traditional ways under the same recall rate; and the the feature extraction time is about 5 to 10 times faster on the CPU and 100 times on the GPU.

Key words: visual simultaneous location and mapping (vSLAM), loop closure detection, convolutional neural network, deep learning, similarity