When a vehicle exits a tunnel, it may face bright sunlight on a clear day, or encounter shadows, wet road surfaces, and dense traffic.
Although the tunnel exit appears to be just an ordinary stretch of road, it concentrates imaging, object recognition, multi-sensor fusion, and vehicle control into a single scenario. For intelligent driving systems, diverse environmental changes test whether cameras can clearly see lane markings, leading vehicles, and obstacles on the road, and whether perception results can continue to support vehicle decision-making.
01. Why Does Sudden Light Change Affect Cameras?
For intelligent driving vehicles, cameras do not record the real world exactly as it is.
Light entering the image sensor must undergo exposure, signal conversion, and image processing to form images used by perception models.
When a vehicle moves from inside a tunnel to a strong light environment, the original exposure settings may no longer be suitable for the new lighting conditions. If exposure time or gain adjustment is not timely, overexposure may occur in the sky, road surface, and vehicle surfaces at the exit. Details such as body contours, lane markings, and even low-profile obstacles that should be visible may be lost.
Conversely, if exposure is limited to preserve details in bright areas, targets in the tunnel interior or surrounding shadows may become too dark.
In addition, camera exposure cannot only aim for overall brightness in the image.
What autonomous driving truly cares about is whether key areas such as roads, vehicles, and pedestrians retain sufficient information.
The fact that the entire image looks clear does not mean every critical target can be reliably recognized. Therefore, automatic exposure control needs to adjust exposure time and gain according to scene changes, balancing brightness adaptation, image noise, and motion blur.
If the exposure time is too long, vehicles and pedestrians may exhibit motion blur in the image; if the exposure time is too short, it may lead to a decrease in the signal-to-noise ratio under insufficient lighting; if gain is increased to improve image brightness, noise may also become more prominent.
At high speeds, exposure control must balance brightness, noise, and motion blur, and cannot rely solely on extending exposure time to improve imaging in dark areas.
02. What Can HDR Solve and What Can It Not?
HDR, or High Dynamic Range imaging, is primarily designed to allow cameras to better record details in both bright and dark areas within the same scene.
For environments with significant contrast between light and shadow, such as tunnel exits, HDR can reduce the risk of detail loss in highlight areas and help improve visibility in shadow parts.
However, HDR does not mean that cameras can see clearly under any lighting conditions. It improves the sensor's ability to record different brightness levels but does not guarantee that all details are completely preserved.
The final effect still depends on the dynamic range of the image sensor, the specific HDR implementation method, the exposure control strategy, and subsequent image processing.
Strong direct sunlight, lens dirt, rain reflections, and motion blur cannot be completely eliminated by HDR alone.
More importantly, imaging optimization and object recognition are not the same thing.
Image processing can try to preserve details as much as possible, but it cannot guarantee that the model will interpret these details correctly. If lane markings are overwhelmed by strong light, or if the contrast between the leading vehicle's outline and the background is low, the perception system may still experience missed detections, localization deviations, or unstable target tracking.
In 2025, the Journal of Mechanical Engineering published a study on cooperative radar and visual perception for smart vehicles facing tunnel scenarios. The study proposed a camera adaptive exposure control method for sudden light changes at tunnel entrances and exits. Meanwhile, addressing the issue of multipath false targets that millimeter-wave radar may encounter inside tunnels, it designed corresponding suppression strategies and conducted tests on a real vehicle platform.
The paper reported that its proposed cooperative perception algorithm improved detection accuracy by 4.8% compared to other models.
It should be noted that this result comes from the specific algorithms and test conditions described in the paper and cannot be directly extrapolated to imply that all mass-produced intelligent driving systems will achieve the same improvement. However, this study indicates that the perception problem at tunnel exits is not merely a matter of camera parameters.
Imaging quality is the foundation, but the system also needs to judge whether the current image and information provided by other sensors are sufficient to support subsequent perception and decision-making.
03. How Does the Perception System Judge Reliability?
After the camera outputs images, the perception model identifies vehicles, pedestrians, lane markings, and other road elements, and estimates information such as target position and speed.
But the model only provides a recognition result, which does not mean the result is equally reliable in all environments.
During sudden light changes, in addition to recognizing road targets, the system can evaluate the reliability of current perception results by combining detection and tracking results from consecutive frames, image quality metrics, and consistency among different sensors.
For example, whether the position of the same leading vehicle shows unreasonable jumps in consecutive images, whether lane marking recognition suddenly interrupts, and whether target detection results frequently appear and disappear.
Information such as the proportion of overexposed areas, dark area details, and degree of blur can also help the system assess the quality of input images.
However, none of these metrics alone can represent the reliability of perception results. A very bright image does not necessarily mean recognition failure, and a high model score for a detection box does not guarantee that the target's position and motion state are completely accurate.
The system needs to combine information such as image quality, continuity of target tracking, and consistency among different sensors to assess whether the current result is sufficiently reliable.
The key here is to ensure that perception reliability assessment influences subsequent driving decisions, rather than just generating a score within the perception module.
When the system finds that lane marking information is unstable or there is significant uncertainty in judging the status of the leading vehicle, it needs to use these results cautiously, avoiding aggressive actions such as lane changes or rapid acceleration when evidence is insufficient.
04. Can Other Sensors Compensate for Brief Visual Deficiencies?
When cameras are affected by lighting, LiDAR and millimeter-wave radar can provide different forms of environmental information.
LiDAR obtains target distance and spatial structure by measuring laser echoes; millimeter-wave radar provides information such as target distance and relative speed. Their measurement mechanisms differ from cameras, so they can provide valuable supplements when vision is affected by lighting.
However, sensor complementarity does not mean that radar can simply replace cameras. The observation capabilities of cameras, LiDAR, and millimeter-wave radar are not identical, and each has its limitations.
As a special environment, tunnels' walls, ceilings, and other structures may cause multipath reflections for millimeter-wave radar, making some echoes correspond to false targets that do not exist at those positions.
LiDAR may also be affected by target reflection characteristics, occlusion, and point cloud sparsity. Therefore, the system cannot unconditionally trust all output results from a sensor just because it is relatively less susceptible to lighting changes.
Truly valuable fusion involves cross-checking observations from different sensors within a unified spatiotemporal relationship.
Does the leading vehicle identified by the camera match the distance and relative speed measured by the radar? Do the obstacles observed by LiDAR roughly align with the positions of visual targets? If there are conflicts in information, the system needs to make judgments based on the current observation quality of each sensor, continuity of target motion, and time synchronization status.
05. Vehicle Driving Strategies Must Also Change When Visibility Is Poor
The ultimate goal of the perception system is not to output a clearer image, but to provide information sufficient to support driving decisions for the vehicle.
If lane marking recognition near the tunnel exit is temporarily unstable, or if the motion state of the target ahead cannot yet be reliably confirmed, the vehicle should not continue to drive aggressively according to strategies used when information is sufficient.
From a system design perspective, more conservative strategies can be adopted.
For instance, avoid initiating unnecessary lane changes when uncertainty is high, control the aggressiveness of acceleration and lateral trajectory adjustments, and leave more ample following and braking margins when conditions permit.
If critical perception information remains persistently insufficient, the system also needs to take corresponding risk mitigation measures according to its capability boundaries. The specific action taken depends on the road conditions, surrounding traffic, vehicle status, and system design at the time.
This demonstrates that even just the common scenario of a tunnel imposes extremely high requirements on intelligent driving systems. The perception module must not only identify the environment as accurately as possible but also promptly convey the reliability of perception results; the planning and control modules need to adjust driving actions based on available information and current risks.
Only when imaging, perception, fusion, planning, and control form an effective collaborative chain can brief camera interference from strong light avoid easily escalating into unstable vehicle behavior.