Lane detection technology is a crucial technology for achieving autonomous driving of vehicles, facilitating the real-time processing of images captured from cameras for autonomous driving. In recent years, researchers have made significant progress in the accuracy of lane detection. However, in practical driving scenarios, the limited computing power of the vehicle's controller may influence the speed and accuracy of lane detection. This paper presents a method of dynamically identifying regions of interest, using the Donkey car obstacle avoidance vehicle for real vehicle experiments. By reducing interference information in the images captured by the camera in real-time, the lane lines are identified more accurately. With a constant number of collected images, the proposed method improves the speed of convolutional neural network (CNN) model training and the accuracy of lane detection during autonomous driving, thereby reducing the occurrences of crossing the lane or leaving the road due to misjudgment of lane lines. The optimized model is about 38.71% shorter than the original model in terms of training time and about 21.67% shorter in terms of trolley autopilot time, while the trolley travels more uniformly at turns and has less sway from side to side. The experimental results demonstrate that the proposed method exhibits excellent detection performance and accuracy.