Abstract
In complex highway scenarios, autonomous driving systems require robust and adaptive decision-making capabilities to ensure safe and efficient driving operations amidst dynamic multi-vehicle interactions. To address the limitations of traditional vector data inputs in high-dimensional feature perception and the excessive redundancy in end-to-end image inputs, this paper proposes a deep reinforcement learning method based on an image-enhanced data framework SAC (IEDF–SAC). This approach introduces an innovative data representation structure that retains critical traffic object features essential for decision-making while mitigating the interference of environmental background and platform device parameters on the policy, thereby significantly enhancing the policy’s portability and generalization ability. In the feature extraction phase, a lightweight MobileNetV3-Small network is employed as the image feature encoder, incorporating inverted residual structures, attention mechanisms, and efficient activation functions. This enables rapid processing of continuous multi-frame images at a 64 × 64 resolution, achieving efficient and real-time feature extraction. The policy network, built on the soft actor-critic (SAC) algorithm, facilitates smooth control of continuous actions and is trained and validated in a self-constructed multi-vehicle dynamic interference highway simulation environment. Extensive experimental results demonstrate that IEDF–SAC outperforms comparative algorithms such as DDPG and baseline SAC across comprehensive metrics including safety, stability, and driving efficiency. Ablation studies further validate the effectiveness of the image-enhanced data framework. This research not only advances the application of continuous-action deep reinforcement learning in autonomous driving scenarios but also provides a scalable technical solution for the real-time deployment of intelligent transportation systems.
Get full access to this article
View all access options for this article.