Spatial Reasoning and Risk Assessment for Autonomous Vehicles on Consumer Electronics Platforms Using a Customized Vision–Language Model With Data Augmentation

Accurate spatial reasoning and risk assessment from monocular video on consumer electronics platforms are prerequisites for safe decision-making in autonomous vehicles, yet general-purpose vision–language model (VLM) remains unreliable at lane-level localization and temporally grounded risk estimation. We present a deployment-oriented, spatially enhanced VLM that learns implicit 3D reconstructive information from monocular sequences via a reconstructive 3D tokenization pipeline and a spatial–visual fusion encoder built on MobileVLM. Concretely, the model fuses 3D reconstructive tokens with CLIP visual tokens through cross-attention to produce 3D-aware visual tokens, enabling lane-level localization and depth-aware reasoning about object orientation and inter-object relations. The unified decoder jointly generates scenario descriptions, object identities and positions, and converts spatial estimates into interpretable multi-level risk scores using a Time-to-Collision (TTC) mapping. Experimental results show that our system outperforms strong VLM baselines for multi-level risk classification on the nuScenes dataset. The generalization in CARLA and CoVLA further corroborates robust spatial reasoning and risk assessment. The findings indicate that coupling a lightweight VLM with monocular 3D cues via cross-attentional spatial–visual fusion yields accurate spatial localization and interpretable, transferable risk estimation that is robust under shift and practical for edge deployment in downstream tasks.