Lane detection is a technique that uses geometric features as an input to the autonomous vehicle to automatically distinguish lane markings. To process the intricate features present in the lane images, traditional computer vision (CV) techniques are typically time-consuming, need more computing resources, and use complex algorithms. To address this problem, this paper presents a deep convolutional neural network (CNN) architecture that prevents the complexities of traditional CV techniques. CNN is regarded as a reasonable method for lane marking prediction, while improved performance requires hyperparameter tuning. To enhance the initial parameter setting of the CNN, an S-Shaped Binary Butterfly Optimization Algorithm (SBBOA) is utilized in this paper. In this way, the relative parameters of CNN are selected for accurate lane marking. To evaluate the performance of the proposed SBBOA-CNN model, extensive experiments are conducted using the TUSimple and CULane datasets. The experimental results obtained show that the proposed approach outperforms other state-of-the-art techniques in terms of classification accuracy, precision, F1-score, and recall. The proposed model also considerably outperforms the CNN in terms of classification accuracy, average elapsed time, and receiver operating characteristics curve measure. This result demonstrates that the SBBOA optimized CNN exhibits higher robustness and stability than CNN.