In order to obtain an improved robustness to the accuracy recall of V-SLAM loop closure detection in complex environments, a local feature matching algorithm, combined with multiple attention mechanisms in graph neural network, has been proposed with an application to the loop closure detection. Firstly, the SuperPoint detector is used to obtain the key points in the image sequence, followed by an input of the extracted feature points into the key point encoder, with its dimension raised to the same as the local descriptor sub-dimension by using a multi-layer perceptron. Then, a more representative local description can be obtained after being repeated 9 times in a multiple attention mechanism network. Next, the SinkHorn algorithm is used to solve the optimal matching matrix in the optimal matching layer, thus obtaining the loop closure detection result by setting the threshold reasonably. Finally, experiments are conducted, alongside with five other loop closure detection benchmark algorithms, on two common datasets of New College and City Centre. The results show that the proposed algorithm is characterized with a higher accuracy and a stronger robustness than other experimental algorithms under a certain recall rate, meeting the requirements of closed-loop detection.