Abstract: To solve the problem that global positioning system（GPS） was easy to fail in urban environment, an autonomous navigation algorithm for intelligent vehicle was proposed by visual location recognition. To identify the global position of vehicle in the GPS-free environment, the discriminative visual features in the image were obtained through the attention model and the fine-grained feature extraction module, and the matching retrieval between the aerial image and the offline satellite image was realized. According to the vehicle position information, the elite ant colony optimization algorithm was used to output the direction of the road branch ahead for the vehicle and perform global path planning. The results show that the fine-grained feature extraction module can extract more discriminative features. The label-smoothed cross-entropy loss function training can be used to achieve effective identification of actual environmental locations, and vehicles can use the proposed algorithm to navigate autonomously in weak GPS environment.