Ensuring sidewalk compliance with the Americans with Disabilities Act (ADA) is essential for pedestrian safety and accessibility. Traditional 2D inspection methods offer high spatial resolution for surface feature detection but lack height information, while existing 3D approaches capture vertical displacements but often miss fine defects such as cracks due to limited resolution. This study introduces a unified 2D–3D computer vision framework that integrates image and geometric information for robust segmentation, labeling, and measurement of sidewalk trip hazards and cracks. The proposed workflow comprises four main stages: (1) a Visual Geometry Grounded Transformer (VGGT) reconstructs detailed 3D geometry from multi-view imagery; (2) a SAM-assisted segmentation network detects potential defects in 2D space; (3) an RGB-D fusion model embeds depth cues into 2D features for accurate trip hazard identification and alignment with 3D point clouds; and (4) quantified geometric metrics enable detailed ADA compliance evaluation. The framework was validated on diverse sidewalk sites with LiDAR and digital-level measurements as ground truth and compared against MASt3R-SLAM as a baseline. Experimental results show that the proposed Pixel-Point Fusion (PPF) framework achieved a recall of 1.000 and an accuracy of 0.905, outperforming both LiDAR and MASt3R-SLAM-based analysis in detecting ADA-noncompliant defects. PPF also successfully captured fine cracks and 0.25-inch trip hazards that other methods failed to detect. Future work will focus on large-scale deployment, enhancing sensitivity to micro-level trip hazards, and enabling real-time detection to support municipal maintenance planning.