Meet Our Client

Our client is a leading Tier 1 global electronics manufacturer providing advanced sensor suites to major automotive OEMs. As the industry moves toward Level 5 autonomy, they are developing sophisticated hardware capable of sensor fusion—combining RGB cameras, Lidar, and thermal imaging to ensure vehicle safety in all environments.

Their Challenge

To achieve full autonomy, standard RGB and Lidar sensors are insufficient, particularly in low-visibility or high-entropy environments. The client faced a significant Reliability Deficit when integrating thermal imaging, as automated models often lack the pre-existing benchmarks to accurately classify heat signatures at high speeds.

The primary technical hurdles included:

• Ontology Conflicts: Significant misalignment between multi-modal data types (Thermal, RGB, and Lidar) made it difficult to establish a single, trusted "ground truth."
  
  
• Distance & Velocity Accuracy: At highway speeds, the system required absolute precision in estimating object velocity and distance to trigger emergency stopping procedures safely.

Resolution Limitations: Identifying a pedestrian at 100+ meters often resulted in only 3-4 data points on a Lidar point cloud, which is insufficient for sensors to classify accurately without expert human context.

Our Solution

CloudFactory developed a Technical Validation Engine that acted as the ground-truth layer for the client’s perception stack. By aligning complex multi-modal ontologies, we provided the high-fidelity oversight required for hardware-level refinement.

• Data Foundation: We designed a specialized thermal-sensor fusion workflow that resolved critical ontology conflicts, ensuring seamless semantic indexing across all sensor types.
  
  
• Model & Orchestration: We implemented an optimized pre-labeling process that utilized model heuristics to handle routine detections, automatically filtering the data stream for high-complexity segments.
  
  
• Oversight Engine: Our domain-trained experts provided the meticulous manual oversight needed to identify pedestrians and obstacles at distances where automated systems remained uncertain, catching the "unknown unknowns" in thermal signatures.
  
  
• Experience: We integrated directly with the client’s engineering environment via SDKs, allowing them to use the validated ground truth to refine their software, firmware, and image processing at the pixel level.

The Results

The partnership allowed the client to improve their software, firmware, and image processing, transforming raw sensor output into a production-ready perception system.

• Extended Safety Margin: By refining detection accuracy, the client achieved a ~20-meter improvement in model detection distance—a critical margin for emergency braking at highway speeds.

• Quality and Operational Throughput: The specialized workflow delivered a 60% efficiency gain over manual-only labeling, enabling the processing of millions of frames with a 1% error rate.

• High-Precision Scaling: We successfully delivered over thousands of sequences with a 100% acceptance rate, validating millions of individual objects for one of the most advanced CV stacks in the market.