Aiming at the problems such as low robustness caused by the changes of indoor lighting conditions in traditional visual SLAM, an optimized visual SLAM method based on image enhancement and neural network is proposed. This method makes improvements on the original ORB-SLAM2 framework by incorporating a RAH-GCNv2 feature point extraction method into its camera tracking thread. The RAH-GCNv2 method performes equalization processing on the RGB channels of the image to adjust the visual information color bias phenomenon, and conductes adaptive enhancement on the HSV channels of the image to adjust the brightness issue. Through the GCNv2 feature point extraction network, uniformly distributed and scattered feature points were obtained, and experimental verification was carried out on public datasets. The experimental results showed that under underexposure and overexposure conditions, the proposed improved method increased the standard deviation of the captured images by 5 times, the entropy value by 50%, and the average gradient of the images by 5 times. After integrating the RAH-GCNv2 feature point extraction method into the ORB-SLAM2 framework, the camera motion trajectory error was reduced by 30% compared with the original ORB-SLAM2 framework, and problems such as pose loss did not occur. The actual test showed that the trajectory drift problem of the original ORB-SLAM2 framework in weakly textured scenes was corrected, and the mapping effect was significantly improved.