As autonomous driving systems evolve, so too must the testing infrastructure that supports them. While much attention is paid to lane detection, object tracking, and general scene understanding, traffic light recognition remains a critical — yet under-tested — aspect of visual autonomy.

随着自动驾驶技术的不断发展，相应的测试基础设施也必须同步升级。尽管当前测试多聚焦于车道检测、物体追踪和场景理解等方面，但红绿灯识别作为视觉感知系统中的关键环节，却仍处于测试不足的状态。

At Fourier Image Lab, we set out to change that.

在研鼎，我们致力于改变这一现状。

A Purpose-Built Simulation System

一套专为红绿灯感知测试设计的仿真系统

We’ve developed a traffic light simulation and testing platform designed from the ground up to evaluate camera perception systems under real-world signal conditions and failure modes.

我们自主研发了一套交通信号灯仿真与测试平台，专门用于在真实信号条件和故障模式下评估摄像头的感知能力。。

The product consists of a matrix of 40 LED, each containing red, yellow, green, and white emitters. The brightness of each individual row is independently controllable, allowing us to reproduce the entire range of lighting scenarios — from faint incandescent bulbs found in legacy intersections to high-intensity directional LED signals.

该产品采用10×4色LED阵列结构，LED阵列集成了红、黄、绿、白四种发光单元。每一颗LED灯珠的亮度均可独立调节，从而模拟从传统低亮度白炽灯到现代高强度定向LED信号灯的全光照场景。

Traffic Light Simulation Array: 40 rows × 4 colors (Red, Yellow, Green, White), fully programmable

红绿灯仿真阵列：10列 × 4 色（红、黄、绿、白），支持编程控制

This granular control is crucial, because while modern HDR sensors may advertise 100+ dB dynamic range using 16- or even 24-bit containers, the data is ultimately tone-mapped to 8-bit before reaching the neural network. That tone mapping process introduces compression artifacts and quantization noise, often causing a loss of critical precision where it matters most — in small, bright objects like traffic lights.

这种高精度控制尤为关键。尽管现代HDR传感器宣称可支持100+dB动态范围，并使用16位甚至24位容器进行数据处理，但这些数据在送入神经网络之前最终仍会被压缩为8位。该色调映射过程会引入压缩伪影与量化噪声，尤其对交通信号灯这类小而高亮的关键目标，极易造成信息丢失。

Saturation, Hue Drift, Flicker, and Signal Failure

饱和、色偏、闪烁与信号失效模拟

In real-world traffic scenarios, one of the most frequent failure modes is spectral channel saturation — where a red or yellow signal overwhelms one color channel but not the others. This can make it difficult or impossible to distinguish between lights of different colors, especially when the image is tone-mapped and passed through ISP pipelines.

在实际交通环境中，最常见的问题之一是光谱通道饱和：红灯或黄灯的信号可能会覆盖其他颜色，导致图像中的颜色识别错误。该问题在图像经过ISP处理后更易放大。

To further challenge perception systems, our platform also allows precise control over flickering parameters, including frequency and duty cycle. This enables testing under regional signal conditions — from low-frequency flickering incandescent bulbs used in parts of the U.S. to high-frequency PWM-modulated LED signals typical in Asia and Europe. By mimicking these flicker styles, we can identify how various sensors and algorithms respond to temporal instability — a major source of missed or ghosted signals.

为了进一步挑战感知系统，我们的测试平台还支持精确控制信号闪烁参数，包括频率与占空比，可模拟从美国地区低频闪烁的白炽灯，到亚洲和欧洲常见的高频 PWM调光LED信号灯。通过模拟不同的闪烁模式，我们可以测试摄像头传感器和算法对时间不稳定信号的响应能力--造成信号丢失或误判现象的主要原因。

Our system monitors:

l Which channels saturate first under varying light intensities

l The hue and saturation values of red vs. yellow signals across the entire brightness range

l Delta thresholds between hue/sat values, triggering alerts when color signals become ambiguous or fail classification standards

l Signal dropout or misinterpretation under flickering conditions, especially in multi-exposure or rolling shutter sensor architectures

我们的系统可监控以下性能：

l 各通道在不同光强下的饱和顺序

l 红与黄信号在不同亮度下的色调与饱和度变化

l 色调/饱和差值阈值，超出时自动触发警告，指出色彩信号分类失败或模糊

l 在多曝光或滚动快门架构下，闪烁引发的图像掉帧或误判

We treat channel saturation and flicker-induced dropout as functional failures, and flag them accordingly. This helps developers not only benchmark their sensor and ISP stack, but also improve resilience in real-world driving environments.

我们将通道饱和与闪烁导致的信号丢失视为功能性失效，并加以标记。这不仅有助于开发者评估其传感器ISP系统，也能显著增强其在真实场景中的适应性。

Image Analysis and Signal Verification

The platform supports both automated and manual workflows for extracting meaningful metrics from camera-captured frames.

图像分析与信号验证功能

平台支持自动与手动结合的图像分析流程，从拍摄图像中提取关键视觉指标：

Automatic Detection

自动检测

automatic detection of the traffic light pattern

自动识别红绿灯位置

The system detects light positions automatically and aligns them with signal geometry.

系统可自动识别灯光位置，并与信号灯的几何结构对齐。

Manual Corner Assistance

手动辅助校准

manual corner detection

In edge cases where automatic detection fails, users can manually annotate the grid corners — the system will interpolate the signal points

当自动检测失败时，用户可以手动标注网格角点，系统会插值生成信号位置。

RGB Overlay Map

RGB覆盖图

RGB values are computed by averaging pixel values

通过像素平均值计算 RGB信号值

Side-by-Side Comparison of two cameras

两款摄像头的对比测试

Naked eye observations

肉眼观察结果

In the left image, color fidelity is preserved across a wide dynamic range — with most of the signal rows retaining accurate hue and separation. Saturation only occurs at the bottom-most row, where white levels clip as expected under extreme intensity.

左图：色彩保真度在较宽的动态范围内得以保持，大多数的色彩信号都能够保留准确的色调和分离度。饱和现象仅在最下方一行，在极端光环境下，白光如预期那样发生了剪切。

In contrast, the right image exhibits early saturation and color breakdown. From the 5th row downward, red, yellow, and green signals begin to blur into each other, and meaningful color distinction is lost. Only the top 3–4 rows appear properly exposed.

相反，右图：过早饱和，颜色分离能力弱。第五列起红黄绿开始混合，颜色识别失败，仅顶部3–4行曝光正常。

Another noticeable artifact is color bleeding, particularly visible on the top (red) and bottom (white) rows. The red signal appears orange-yellow, and the white signal exhibits an unexpected greenish tint. While the exact cause of this distortion is unclear, it may stem from local tone mapping or histogram equalization techniques (such as CLAHE) that operate non-uniformly across image regions — potentially skewing channel balance and local contrast.

还观察到颜色漂移：顶部红灯偏橙黄，底部白灯出现绿色偏移。这种失真可能来源于ISP中的局部 tone-mapping或直方图均衡算法（如CLAHE），其对不同区域的处理不一致，导致通道平衡失衡。

These results underscore the importance of controlled signal testing under known brightness and flicker conditions. The visual inconsistencies between sensors highlight how critical both sensor behavior and ISP tone mapping are in preserving safety-critical color information.

这些结果强调了在已知亮度和闪烁条件下进行受控信号测试的重要性。不同传感器之间的可视化差异，凸显了传感器行为和ISP色调映射在保持关键安全色彩信息方面的重要性。

Quantitative Analysis: RGB Channel Response to Signal Intensity

量化分析：RGB通道对信号强度的响应

To complement the visual comparison, we analyzed the RGB values of red and yellow signals across increasing intensity levels (rows 1–10) from each camera. These values were obtained by averaging pixel intensities from the center of each illuminated signal.

为了补充上述视觉对比，我们分析了两台摄像头在不同亮度等级（第1至第10列）下红色和黄色信号的RGB值。这些数值通过对每个点亮信号中心区域的像素亮度进行平均计算而得出。

Left Camera - better performer

左侧摄像头 – 表现更佳

Left Camera – RGB values for Red and Yellow signals across rows 1 to 10

左侧摄像头 – 红色与黄色信号在第1至第10列的RGB值

For the red signal, all three color channels (R, G, B) show gradual progression with saturation beginning only at row 7 — which corresponds with the white-out observed in the last row of the photo. This indicates a healthy dynamic response.

红光信号下，RGB通道逐渐上升，到第7列才开始饱和（对应照片中最下方的白溢出），显示出良好的动态响应。

For the yellow signal, however, the green channel saturates first, followed by blue, and finally red. This staggered saturation is a critical insight: when viewing cropped sections of the red and yellow rows from this camera.

黄光信号下，绿通道首先饱和，其次是蓝、再是红。这种交错式的饱和现象是一个关键的发现：当观察该相机拍摄的红色和黄色的裁剪区域时尤为明显。

top row is red that becomes yellow, than white

顶部为红色信号，逐渐变为黄色，然后为白色

We can observe the red signal temporarily adopting a yellow/orange hue before becoming fully saturated. This shift can fool a classifier into reading red as yellow — a safety-critical misclassification scenario.

可以观察到红色信号在完全饱和前会暂时呈现出黄色或橙色的色调。这种色彩偏移可能会误导算法将红色识别为黄色，从而导致严重的安全性误判。

Right Camera (underperformer)

右侧摄像头（表现不佳）

Right Camera – RGB values for Red and Yellow signals across rows 1 to 10

右侧摄像头 – 红色与黄色信号在第1至第10列的RGB值

Here, the data tells a more alarming story. For the red signal, all RGB channels saturate fully by row 3, offering no gradient or usable color discrimination beyond that point.

这组数据传递了一个更令人担忧的讯息。对于红色信号，所有RGB通道在第3行前已全部饱和，之后无法再提供任何色彩梯度或有效的颜色识别。

The yellow signal is even more problematic: green is saturated from the very first row, blue reaches max level by row 3, red climbs and saturates around row 5.

黄色信号问题更严重：绿色通道从第一行开始就已饱和，蓝色在第3列达到最大值，红色在第5列左右饱和。

This kind of response indicates that the camera is not capturing the color signal faithfully at all, and that color balancing or tone-mapping inside the ISP may be distorting or flattening the output. In fact, none of the signals were represented correctly — making this camera (an actual automotive-grade sensor with auto-exposure enabled) highly unsuitable for traffic light detection.

这种响应表明，该摄像头根本无法真实还原色彩信号，ISP的曝光设置无法重现交通灯色彩。事实上，所有信号的颜色表现都不正确——这使得这台摄像头（即便是具备自动曝光功能的车规级传感器）在交通信号灯检测应用中极其不适用。

These plots, combined with our structured simulator and visualization tools, highlight the urgency for rigorous camera qualification in perception systems. Even among sensors marketed as “HDR-ready,” internal ISP design choices, tone-mapping, and exposure logic can lead to catastrophic misinterpretations of safety-critical cues like traffic signals.

这些图表结合我们的结构化模拟平台和可视化工具，凸显了对视觉感知系统中摄像头进行严格认证的紧迫性。即使是那些号称支持“HDR”的传感器，在其ISP设计、色调映射和曝光控制策略上，仍可能导致对关键信号的严重误读。