Thermopile sensor arrays provide a sufficient counterbalance between person detection and localization while preserving privacy through low resolution. The latter is especially important in the context of smart building automation applications. Current research has shown that there are two machine learning-based algorithms that are particularly prominent for general object detection: You Only Look Once (YOLOv5) and Detection Transformer (DETR). Over the course of this paper, both algorithms are adapted to localize people in 32 × 32-pixel thermal array images. The drawbacks in precision due to the sparse amount of labeled data were counteracted with a novel generative image generator (IIG). This generator creates synthetic thermal frames from the sparse amount of available labeled data. Multiple robustness tests were performed during the evaluation process to determine the overall usability of the aforementioned algorithms as well as the advantage of the image generator. Both algorithms provide a high mean average precision (mAP) exceeding 98%. They also prove to be robust against disturbances of warm air streams, sun radiation, the replacement of the sensor with an equal type sensor, new persons, cold objects, movements along the image frame border and people standing still. However, the precision decreases for persons wearing thick layers of clothes, such as winter clothing, or in scenarios where the number of present persons exceeds the number of persons the algorithm was trained on. In summary, both algorithms are suitable for detection and localization purposes, although YOLOv5m has the advantage in real-time image processing capabilities, accompanied by a smaller model size and slightly higher precision.

Perception-oriented evaluation of automotive headlamps still relies mainly on human-vision photometric criteria, although forward-facing cameras are increasingly safety-critical sensing elements for night driving. This paper benchmarks 16 measured production headlamp light distributions with a simulation chain that combines headlamp spectra and beam patterns, diffuse scene reflection, an imaging-transfer model, and an EMVA-based camera model. The quantitative chain maps scene radiance to sensor-domain signal-to-noise ratio, derives task-specific required signal-to-noise curves from a six-network object-recognition ensemble, and aggregates local threshold satisfaction as region-of-interest coverage across three target reflectances and five driving speeds using WLTP moving-time weights. For the baseline RGB camera, WLTP-weighted coverage ranges from 18.95% to 53.48% across the evaluated light distributions, corresponding to a factor of 2.82 between the weakest and strongest distribution. The camera-parameter sweeps show that favorable beam placement can deliver comparable benchmark coverage with roughly 60% smaller pixel pitch than the weakest distribution, corresponding to an 84% reduction in pixel area, or at materially shorter exposure times. The WLTP-weighted coverage score correlates positively with the established Headlamp Safety Performance Rating, with Pearson $r=0.68$ for the RGB configuration, indicating partial alignment between human-centric and camera-centric illumination needs while confirming that the metrics are not interchangeable. The results identify headlamp light distribution as a primary design parameter for nighttime camera perception and provide a quantitative basis for co-design of automotive lighting and camera-based systems.