Optimization of Lightweight Deep Neural Network Models for Steering Angle Prediction

In this paper, we focus on achieving precision improvement in steering angle prediction with a lightweight deep neural network model. We trained an end-to-end deep neural network model to achieve three-task autonomous driving for steering angle prediction. The design and optimization work of the model is discussed in this paper. Our proposed CNN-BiLSTM model combines CNN and bidirectional LSTM. Meanwhile, we improved the training and testing performance of the model by fine-tuning the hyperparameters. The optimized model has low training loss and good stability. In addition, we proposed a visualization method for model prediction to optimize the evaluation of the model. The prediction visualization method tagged the predicted steering angle of the sample data automatically. The CNN-BiLSTM model outperforms existing models in steering angle prediction with 97.6% F1.