An Intelligent Navigation Control Approach for Autonomous Unmanned Vehicles via Deep Learning-Enhanced Visual SLAM Framework

For autonomous unmanned vehicles (AUVs), navigation control is an essential concern in academia. Conventionally, existing research works dealt with this issue by resorting to the “Simultaneous Localization and Mapping” (SLAM) technology. Although many SLAM-based approaches had been proposed in recent years, they could not mostly perceive semantic characteristics from dynamic visual scenarios. To deal with this challenge, this paper proposes a navigation control approach for AUVs via a deep learning-enhanced visual SLAM framework. Firstly, the commonly used coordinate system in the motion of AUVs is analyzed, and a mathematical description of the circular arc motion of AUVs is formulated. Then, an up-convex curve model is adopted to realize high-accuracy detection of the bilateral lane lines. On the foundation, a yaw angle guidance-based imitation learning framework is utilized to realize navigation control. This well facilitates the analyzing the causal relationship between scenarios and decisions. Some experiments are conducted by simulating real-world urban road scenes, to testify the efficiency of the navigation control for AUVs. The results show that navigation accuracy can be improved by about 5%.