Abstract:

Existing research on autonomous driving scene representation tends to encode sensor data into high-dimensional features to retain maximum raw information. However, this approach introduces substantial background noise irrelevant to decision-making and incurs a heavy computational burden, especially when fusing historical features to extract temporal dynamics. To address these challenges, this paper proposes DSVec (Dynamic-Static Vectorization), a novel vectorized scene representation method based on dynamic-static feature fusion. First, complex traffic scenes are abstracted into a series of structured dynamic and static vector elements. On this basis, a dynamic-static feature fusion network is designed, which leverages Variational Auto-Encoders (VAE) to extract low-dimensional static features from semantic Bird's-Eye-View (BEV) maps. These are combined with historical trajectory features of dynamic obstacles extracted via graph-based structures, using Temporal Convolutional Networks (TCNs) to achieve precise spatiotemporal alignment of heterogeneous features. Subsequently, a Transformer decoder augmented with a category-level masking mechanism is introduced to eliminate redundant information via attention mechanisms, realizing both deep fusion of dynamic-static features and independent decoupling reconstruction of various vectorized elements. Finally, a deep reinforcement learning decision-making model based on the Soft Actor-Critic (SAC) algorithm is constructed using the compact vectorized state space generated by DSVec, and systematic validation is conducted through end-to-end closed-loop simulations on the CARLA high-fidelity platform. Experimental results demonstrate that the proposed method achieves high accuracy in reconstructing both dynamic and static elements. Compared with traditional rasterization-based representation methods, DSVec significantly enhances the environmental adaptability, decision-making safety, and computational efficiency of autonomous vehicles in complex dynamic scenarios such as roundabouts and unprotected intersections.

autonomous driving; vectorization scene representation; trajectory predicion