Introduction Computer vision in autonomous vehicles is not one application but ten distinct subsystems, each with its own production constraints, datasets, accuracy budgets, and failure modes. The popular framing treats AV CV as a single end-to-end model that “sees the road”; the production reality is a pipeline of specialised perception components — lane detection, pedestrian tracking, sign recognition, depth estimation, sensor fusion, drowsiness monitoring, parking assist, blind-spot detection, traffic-light recognition, and road-condition assessment — each engineered against safety budgets that allow no marketing tolerance. See computer vision for the broader subdomain this article maps onto. The naive read is that one model architecture scales to all ten. The expert read is that the ten subsystems have different latency budgets, different sensor inputs, different ground-truth pipelines, and different failure consequences — and the right architecture for each is the one whose constraints match the subsystem, not the one that fits the team’s GPU. What this means in practice AV CV is ten distinct subsystems, each scoped against its own safety and latency budget. L2 ADAS and L4 autonomy share components but ship different validation evidence. Occlusion, weather, and long-tail events remain the unsolved problems in 2026. Sensor fusion (camera + radar + lidar) changes the architecture; camera-only is bounded. What are the production-validated CV applications in autonomous vehicles today (lane, sign, pedestrian, depth)? Ten subsystems have crossed from research demo to production-validated deployment. Lane detection: segmentation networks against highway and urban datasets, used in L2 lane-keeping and L4 path planning. Pedestrian detection and tracking: object detection plus multi-object tracking, with cyclist as a distinct class, validated against KITTI/Waymo/nuScenes pedestrian splits. Traffic-sign recognition: detection plus classification against regional sign datasets, with the practical constraint that signs vary by jurisdiction. Depth estimation: stereo-camera and monocular networks, with stereo dominant where the baseline allows and monocular accepted where it does not. Sensor fusion: camera-radar fusion for L2, camera-radar-lidar for L4, fusing detections at object or sensor level. Drowsiness and driver monitoring: in-cabin face and gaze tracking, validated in L2 attention-supervision systems. Parking assist: surround-view stitching plus obstacle detection in low-speed regimes. Blind-spot detection: side-camera or radar-fused detection of vehicles in adjacent lanes. Traffic-light recognition: detection plus state classification, validated against the wider colour and shape variance than research datasets capture. Road-condition assessment: surface classification for traction, wet/snow/dry, used in adaptive control. Each subsystem ships with its own validation evidence; combining them into a stack is a separate engineering exercise. How do CV stacks differ between Level 2 ADAS and Level 4 autonomous deployments? L2 ADAS stacks are designed around the driver-in-the-loop assumption. The CV components target driver-assist accuracy bands: lane-keeping that handles 90% of highway conditions with the driver expected to take over for the rest, AEB that triggers at conservative thresholds because false positives erode trust faster than missed positives erode safety, driver monitoring that backstops attention. Sensor stacks are typically camera plus radar; lidar is rare. Validation evidence targets the L2 regulatory frame and the consumer use cases. L4 autonomous stacks are designed around the no-driver assumption. The CV components target full-redundancy accuracy bands — pedestrian detection that handles occlusion and degraded weather without driver fallback, sign and traffic-light recognition that handles regional variance without operator override, depth estimation that supports planning without driver judgment. Sensor stacks are camera plus radar plus lidar, often with redundant cameras. Validation evidence targets ODD-specific (operational design domain) safety cases and includes formal scenario coverage that L2 does not require. The stacks share components in name only; the engineering envelope, the validation depth, and the deployment unit cost differ by an order of magnitude. Which CV problems remain unsolved in 2026 AV pipelines — occlusion, weather, rare events? Three problem classes remain unsolved in the precise sense that no production system handles them at the accuracy required for unrestricted deployment. Occlusion: a pedestrian partially behind a parked vehicle, a cyclist behind a bus, a child between cars — detection and trajectory prediction degrade sharply with occlusion, and the training data is structurally under-represented because occluded objects are hard to label. Adverse weather: heavy rain, snow, fog, low-angle sun degrade camera performance and degrade lidar in different ways; sensor fusion mitigates but does not eliminate. Rare events (long tail): construction zones with non-standard signage, emergency vehicles in non-standard configurations, road debris of unfamiliar shape, animals — the events that dominate disengagement reports for L4 systems are the events the training data did not cover. Each class has active research and incremental progress; none has a closed-form solution in 2026. The honest framing for buyers: ask vendors how their system behaves at the boundary of each problem class, not whether the system handles them, and read disengagement reports rather than marketing accuracy numbers. What datasets and benchmarks drive AV CV progress (KITTI, Waymo, nuScenes), and where do they fall short? KITTI established the AV CV benchmark template and remains in use for academic comparison, though its scale and diversity are now exceeded by newer datasets. Waymo Open Dataset and nuScenes are the production-relevant benchmarks: multi-modal (camera + lidar + radar), multi-city, with sufficient scale for modern model evaluation. Argoverse and ONCE add geographic diversity. The benchmarks drive progress on the metrics they measure: 3D detection accuracy on labeled objects, tracking consistency on labeled sequences, segmentation on labeled scenes. The shortfall: the benchmarks under-represent exactly the problem classes that determine production safety. Occlusion is sparsely labeled, adverse weather is sparsely sampled, rare events are by definition rare and so under-counted in any finite sample. The benchmark-best model is not the production-safest model; the gap is most visible in disengagement metrics from operating fleets that the benchmarks do not capture. Vendors that quote benchmark accuracy without disengagement evidence are selling the easier number. The diligent buyer reads both. How does sensor fusion (camera + radar + lidar) change the CV architecture versus camera-only stacks? Camera-only stacks (Tesla-style) push the architecture toward end-to-end vision: a large model consumes multi-camera input and emits the planning-relevant representation, betting that scale on data and compute compensates for the absent ranging sensors. The architecture is uniform and the deployment cost per vehicle is low; the validation surface is large because the model is monolithic, and the failure modes are correlated across cameras (rain affects all cameras together). Camera + radar + lidar stacks (Waymo-style, most L4) push the architecture toward sensor fusion: per-sensor perception components produce detections that are fused at object level or feature level, with the fusion layer responsible for resolving disagreement and exploiting complementary failure modes (lidar handles low-light where camera fails, radar handles fog where lidar degrades). The architecture is modular, the deployment cost per vehicle is high (lidar dominates the BOM), the validation surface is smaller per component and larger at the fusion interface, and the failure modes are decorrelated across sensors. The architecture choice cascades through dataset strategy, validation strategy, and unit economics; it is not a sensor decision but an organisational one. What is the role of classical OpenCV-style CV in a modern AV stack alongside end-to-end deep learning? Classical CV is not obsolete in modern AV stacks; it is concentrated in roles where its determinism and low compute cost are advantages. Camera calibration, image-rectification, lens-distortion correction, and stereo-block-matching for short-range depth remain classical. Sensor-time synchronisation, geometric verification of detections (does the lidar point cloud agree with the camera detection at the predicted 3D position), and post-processing of deep-learning outputs (non-maximum suppression, temporal smoothing, multi-frame association in regimes where Kalman-style filtering outperforms learned tracking) are largely classical. The pattern: deep learning dominates where the relationship between input and label is too complex to specify analytically (object detection, semantic segmentation, monocular depth); classical CV dominates where the relationship is well-understood geometrically (calibration, projection, fusion) or where determinism is required for safety certification. AV stacks that throw out classical CV produce architectures that are harder to certify and slower to debug; stacks that ignore deep learning produce systems that cannot handle perception variance. The production-correct stack is layered, with each technique in the role its constraints suit. Limitations that remained The ten-subsystem framing is the production view as of 2026; the field is moving fast on end-to-end architectures that aim to collapse subsystems into single models, and the framing will shift as that work matures. Per-subsystem accuracy numbers are deployment-specific and vary across vendors, sensor configurations, and operating design domains; treat any quoted figure as conditional on those parameters. Disengagement reporting, the most operationally meaningful safety metric, is not standardised across jurisdictions and is selectively published; buyers should request operating data under NDA rather than relying on public reports. Sensor fusion architectures are evolving rapidly; the camera-radar-lidar baseline of 2026 will not be the 2028 baseline as 4D radar and high-resolution low-cost lidar shift the trade-offs. How TechnoLynx Can Help TechnoLynx works on AV CV subsystem engineering and validation — from per-subsystem dataset strategy and model selection through sensor-fusion architecture and the integration layer that turns ten subsystems into a stack. If your team is scoping or auditing AV CV components for L2 or L4 deployment, contact us. Image credits: Freepik
