The split flying car is composed of a flight module, a passenger capsule and an intelligent chassis module. The autonomous docking between these modules enables the split flying car to switch between flight mode and driving mode. The positioning of the aircraft bracket junction point is crucial for determining the desired position of the flight module. However, the complex and variable takeoff and landing environments and the limited computing power of edge computing platforms pose significant challenges to the perception task. To address these issues, we propose a lightweight network-based aircraft bracket detection model that meets real-time requirements in docking scenarios. Firstly, we use the inverse perspective mapping stitched bird's eye view as input to obtain the junction point coordinates of the aircraft bracket through the junction point detector. Then the position information of the bracket is obtained by eliminating the mis-detected junction points and reasoning out the missed junction points based on the a priori information of the aircraft bracket. To facilitate vision-based aircraft bracket detection research, a dataset is established, which is the first publicly available dataset in this research field, collecting 4,631 bird's eye views in different environments. The proposed method can achieve FPS of 35.79 and average precision of 0.915 in the Jetson AGX Xavier edge computing platform. The proposed method can also achieve competitive results when applied in parking slot detection with at least 2 ​× ​faster inference speed.

Autonomous flying vehicles (AFVs) are promising future vehicles, which have high obstacle avoidance ability. To plan a feasible path in a wide range of cross-country environments for the AFV, a triggered forward optimal rapidly-exploring random tree (TF-RRT*) method is proposed. Firstly, an improved sampling and tree growth mechanism is built. Sampling and tree growth are allowed only in the forward region close to the target point, which significantly improves the planning speed; Secondly, the driving modes (ground-driving mode or air-driving mode) of the AFV are added to the sampling process as a planned state for uniform planning the driving path and driving mode; Thirdly, according to the dynamics and energy consumption models of the AFV, comprehensive indicators with energy consumption and efficiency are established for path optimal procedures, so as to select driving mode and plan driving path reasonably according to the demand. The proposed method is verified by simulations with an actual cross-country environment. Results show that the computation time is decreased by 71.08% compared with Informed-RRT* algorithm, and the path length of the proposed method decreased by 13.01% compared with RRT*-Connect algorithm.