Overview
Advanced driver assistance has become a common automotive feature, enhancing vehicle and driver perception, reducing driver workload, and improving driving safety. Cameras based on CMOS image sensors are one of the primary tools for environmental perception in these systems.
Figure 1: Photoelectric effect
Sensor as an optical memory
Architecturally, a CMOS image sensor resembles a memory: it contains many storage elements with row and column addressing. The difference is that while a memory stores electrically written data, the sensor stores charge written by visible or near-infrared light.
The collected charge in the sensor consists of two parts: the desired charge produced by incident light, which forms the useful signal, and unwanted charge produced by various interference sources, generally referred to as noise.
We prefer more useful information and less interference. A common metric for quantifying image noise is the signal-to-noise ratio (SNR), the ratio of signal to noise. Higher SNR means lower relative noise and better image quality. SNR can be expressed as a linear ratio or in decibels (dB).
CMOS sensor datasheets usually provide an SNR parameter, as shown below.
SNR depends on illumination and operating point
A common question is whether a sensor with 46 dB SNR performs better in low light than one with 43 dB. The answer is: not necessarily. SNR is not a single fixed point but a function of illumination. As illumination changes, SNR traces a curve. Datasheet SNR values are often the maximum SNR, corresponding to very bright illumination, i.e., the SNRMax point on the right-hand side of the SNR curve. For low-light performance, the SNR value in the lower-left region of the curve is the relevant metric.
Comparing only the maximum SNR can be misleading. Users should evaluate SNR at specific exposure conditions that represent the application. For example, evaluate the SNR value at a chosen exposure, or fix an SNR target (for instance SNR=5) and compare the exposure required to reach that SNR. A smaller required exposure to reach the same SNR indicates better low-light performance.
Figure 2: SNR curve
HDR and non-monotonic SNR behavior
Automotive sensors must cover all-day lighting conditions and therefore require high dynamic range (HDR). A common HDR method changes sensor sensitivity to sample different brightness levels, maps multiple exposure frames into a standardized linear data space, and then selects pixels from frames with different sensitivities to construct a complete image.
Changing sensitivity shifts the SNR curve in the coordinate system. The HDR image SNR becomes a composite fit of multiple SNR curves. This composite SNR is not necessarily monotonic: besides low-light SNR being small, local minima can appear in high-brightness regions. If the operating point falls into an SNR drop region, image noise can worsen even under bright conditions. This leads to the counterintuitive outcome that image quality may improve with increasing brightness and then suddenly degrade. Therefore, unlike traditional linear sensors, automotive wide-dynamic sensors require evaluation of local SNR minima in bright conditions.
Figure 3: HDR image SNR
Figure 4: SNR and noise
Temperature, exposure, and analog gain effects
Automotive sensors are analog devices. Before the ADC, signal and noise are stored as charge. Dark current also accumulates charge; its generation rate is roughly proportional to exposure time and depends exponentially on temperature. Therefore temperature, exposure time, and analog gain affect dark-state noise. SNR is simultaneously a function of temperature, exposure time, and analog gain, forming a multidimensional family of curves.
For example, Figure 5 shows how the exposure required to achieve SNR=10 changes with internal sensor temperature. Different sensors have different sensitivity to temperature drift; the exposure curves for the two example sensors cross at 60°C, which implies that SNR-based conclusions can be opposite when evaluated at 25°C versus 80°C.
Figure 5: Effect of internal wafer temperature
In typical automotive camera operating lifecycles, the sensor internal temperature often exceeds 40°C for more than 88% of the time, exceeds 60°C for more than 80% of the time, and exceeds 80°C for more than 65% of the time.
Temperature drift can further depress SNR curves. As indicated by the red dashed curve in Figure 6, elevated temperature reduces SNR in both low-light regions and the local minima in bright regions.
Figure 6: Temperature impact on HDR SNR
Standards and evaluation
The industry commonly uses the EMVA1288 standard from the European machine vision association to evaluate automotive image sensor SNR. EMVA1288 defines SNR based on a traditional monotonic linear sensor model, which does not fully describe the SNR characteristics of automotive wide-dynamic sensors.
The IEEE is working on a new image-quality test standard, P2020, to define automotive imaging metrics. Participants in this effort are contributing to drafting image-noise metrics, including SNR, that better reflect wide-dynamic-range and temperature-dependent behaviors.
Conclusion
SNR is a key image-quality metric for automotive CMOS image sensors. It exhibits non-monotonic behavior and is influenced by multiple factors in automotive environments. Comprehensive and objective SNR evaluation across relevant illumination levels, exposures, gain settings, and temperatures is required to accurately characterize sensor performance and to guide automotive imaging product development.
