Deep Learning‐Driven Steering Angle Prediction and Scene Understanding via Harmonic Carpet Weaver Optimization

ABSTRACT Steering angle prediction refers to forecasting the steering angle as early as possible in advance to ensure smooth and precise vehicle control. Therefore, this article introduces a novel approach, called Quasi Recurrent Neural Network with Harmonic Carpet Weaver Optimization (QRNN_HarCWO), for predicting the steering angles. Primarily, the input video is assimilated from the database, and later, the video is split into multiple frames. Then, the pre‐processing procedure is done by the Gaussian filter, which is exploited to denoise the video frame. Subsequently, the pre‐processed output is fed into the scene understanding module, where semantic segmentation is performed by exploiting BlitzNet with the BiTopK loss function. Here, the BiTopK loss function is a new loss function formulated using binary cross‐entropy and the TopK loss functions. Further, Harmonic Carpet Weaver Optimization (HarCWO) is utilized to train the BlitzNet. Finally, the Quasi‐Recurrent Neural Network with HarCWO (QRNN_HarCWO) is used for performing steering angle prediction, based on which appropriate control action is taken. Moreover, the presented approach gained the minimum values of Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Percent Bias (PBIAS) of 0.0368, 0.1918, 0.1766, and 0.0589. Also, the maximum values of 92.877%, 94.887%, 90.766%, and 88.876% are attained by the QRNN_HarCWO model for accuracy, True Positive Rate (TPR), Positive Predictive Value (PPV), and Negative Predictive Value (NPV) metrics. While the results demonstrate high accuracy, potential real‐world challenges such as sensor noise and hardware limitations may affect the deployment of the model in real settings, which need to be addressed in future work.