In today’s drone positioning methods, reliance on built-in GPS chips and Inertial Measurement Units (IMUs) is prevalent. However, these devices can be costly for general-purpose drones. In our research, we employ the Webots application to simulate the flight routes of typical drones and generate datasets by capturing the differences in angle and distance between consecutive image frames. Subsequently, we use these datasets to train a regression Convolutional Neural Network (CNN) model. This model predicts changes in rotation angles, thereby enhancing flight accuracy during route flights. Unlike the single-camera visual positioning methods, which are based on visual Simultaneous Location And Mapping techniques (vSLAM), our approach focuses on assisting positioning rather than full location. The information generated by the ML model encompasses only a portion of the 3D spatial features, specifically the differences in viewing angles, rather than the entire 3D structure. This approach symbolizes that if we do not extract a significant amount of information (3D space) from limited data (2D images), it is still possible to accurately convert information from 2D to 3D. To reduce the computational load on the drone’s computing platform, we tune the ML model using a lite convolutional layer. Ultimately, we identify the best regression model through a comparison of various model parameters. This model significantly improves flight accuracy during route flights, even in complex environments. Additionally, we evaluate the ML model’s performance on the drone using the roofline performance model and operations.