Many people assume that as long as autonomous driving can recognize vehicles, pedestrians, and traffic lights, it already possesses basic road perception capabilities. However, once actually driving, one realizes that seeing objects on the road and knowing where the vehicle should go are two entirely different things. If lane markings on a road are clear, the vehicle can easily obtain explicit lane references; but once encountering faded markings, snow cover, or construction detours, the originally clear lane markings may suddenly become useless. This raises a question: without lane markings, can autonomous driving still determine where the road is?
01. Why Are Lane Markings So Important for Autonomous Driving?
Lane markings can provide autonomous driving with a relatively stable lane reference. After recognizing lane markings through sensors such as cameras, the vehicle can further estimate the left and right boundaries of the lane, the direction of extension, and the current lateral position of the vehicle within the lane. This information directly participates in the vehicle's trajectory planning and lateral control. However, lane markings are merely markers on the road, not the road itself. Not all roads have perfect lane markings. Markings may become blurred due to long-term wear and tear, or they may be obscured by puddles, snow, or mud.
In 2025, a team from the University of South Florida and Fuzhou University published "Empirical Performance Evaluation of Lane Keeping Assist on Modern Production Vehicles" on arXiv. Based on the OpenLKA dataset (389.07 hours of real-world driving data, 1,166 routes, and over 50 mass-produced vehicle models), they conducted a statistical analysis of real-world road conditions. The results revealed that faded lane markings, low contrast of road markings, and sharp curves are the primary individual factors leading to LKA (Lane Keeping Assist) system failures. Moreover, these factors often occur in combination, significantly increasing the probability of failure. A research team from Tongji University used an onboard holographic information collection system to conduct multiple round-trip tests covering a total of 3,200 kilometers on the Guidu Expressway in Guizhou Province. They analyzed the impact of 11 feature variables on lane marking perception failure based on the CatBoost model.
The research conclusions indicate that under conditions such as a larger maximum downstream curvature, a greater rate of slope change, exit and entrance sections, longitudinal and transverse deceleration markings on the road surface, adverse weather, and higher ego-vehicle speeds, the probability of lane marking perception failure significantly increases. Another study on autonomous driving failures covering national highways, local roads, and urban roads in South Korea (with a total driving distance of 36,244 kilometers, including 14,032 kilometers on national highways) showed that on national highways, lane departure accounted for 20.4% of autonomous driving failure events, making it the second highest failure type second only to system disengagement (57.8%); while on local roads, the proportion of lane departure was only 1.7%. If autonomous driving could only rely on lane markings, the aforementioned scenarios would become obvious capability shortcomings. However, autonomous driving cannot rely solely on lane markings. So, when lane markings fail to provide reliable references, what other methods can vehicles use to determine the road?
02. How Does Autonomous Driving Determine Where the Road Is Without Lane Markings?
After lane markings disappear, the vehicle does not suddenly become blind to the road. Sensors such as cameras and LiDAR can still acquire environmental information around the road. What the system needs to do is utilize this information to determine the road boundaries and the drivable area. The most intuitive reference for judgment is the road boundary. Objects such as curbs, guardrails, central reservations, and green belts can all help the system determine roughly where the road extends. The feasibility of this approach has been quantitatively verified. A study on road boundary extraction based on LiDAR point clouds showed that the method of extracting road boundaries using point cloud spatial features achieved over 90% in key metrics such as completeness, accuracy, and F-Measure. A study from Chongqing University of Technology used a 16-line LiDAR to collect point cloud data from three different unstructured road scenarios, achieving a successful detection rate of 90.04%.
Another study on road edge recognition algorithms based on multi-line LiDAR performed matching and filtering according to the geometric features of the road edge and the 3D point cloud distribution features. Experiments showed that the recognition accuracy was greater than 95% in both straight and curved road scenarios, and the processing time per frame was less than 15 milliseconds, demonstrating good accuracy and real-time performance. However, merely finding the road boundary is not enough. In real-world roads, abnormal situations often occur, such as ongoing construction ahead, where originally passable areas are occupied by traffic cones and barricades. The system also needs to determine which spaces are currently safe to pass through, which is the problem solved by the drivable area. It focuses not on where the lane markings are, but on where the vehicle can drive in the current environment.
On roads without clear lane markings, the system can comprehensively judge the drivable space by integrating information such as road boundaries, obstacles, and construction facilities. A drivable area segmentation model named RDAS-MODEL achieved an mAP@50 of 95.2% and an mAP@50:95 of 77.2% on the BDD100K dataset, with an inference speed reaching 102 frames per second, significantly outperforming contemporary advanced methods. The movement of other vehicles ahead can also provide auxiliary information. Mobileye's REM (Road Experience Management) crowd-sourced mapping solution is a practice based on this very concept. Its core logic is not to rely on a single vehicle's real-time recognition of lane markings for localization, but to build a global 3D map through road information anonymously collected by millions of sold vehicles.
As of now, over 8 million vehicles have contributed approximately 34 billion miles of driving data to this platform, and the REM map covers over 95% of public roads in the US and Europe. Stellantis has announced that starting in 2027, it will equip this solution in its brand vehicles to achieve intelligent lane keeping on unmarked roads. This is a typical ODD (Operational Design Domain) restricted scenario, relying on the real-time performance of high-precision crowd-sourced maps to ensure safety redundancy. Of course, following the vehicle ahead does not mean the problem is solved. The leading vehicle itself may make incorrect maneuvers, so the trajectories of other vehicles can only serve as part of the environmental information and cannot be the sole basis. In March 2018, an Uber autonomous driving test vehicle was involved in a fatal collision in Tempe, Arizona.
The accident investigation revealed that the vehicle had detected the pedestrian about 6 seconds before the collision, but the software repeatedly switched between classifying the object as a pedestrian and as something else, leading to decision-making delays and a failure of the system to react in time. This accident directly led to Uber suspending its autonomous driving tests across the US, becoming a landmark event in the history of autonomous driving development. It demonstrates that when the perception system cannot form a stable judgment among multiple information sources, even if each source individually detects something, the system may still make incorrect decisions. Besides using crowd-sourced trajectories as an auxiliary, another approach is to enhance single-vehicle perception, which is the route Tesla has leaned towards.
Its Occupancy Network discretizes 3D space into small cells, using 0/1 to indicate whether each cell is occupied, thereby modeling the passable areas in 3D space. After introducing the Occupancy Network, the FSD (Full Self-Driving) system's perception capabilities for irregular objects and spaces have been significantly improved. This technical route based on 3D space occupancy modeling reduces the reliance on lane markings at the architectural level. In its latest patent, Tesla's system can identify details such as curbs, parking lines, disabled parking symbols, and even fire lanes, with accuracy reaching three times that of current systems, while processing power consumption is reduced by 90%.
03. Final Words
Thus, it can be seen that whether it is crowd-sourced maps or single-vehicle perception, autonomous driving perception is evolving from single-target recognition to spatial understanding. What the vehicle truly needs to obtain is not just that there is a car here or a pedestrian there, but also where the road is, where the boundaries are, which areas are passable, and how these areas are connected. Lane markings are indeed important, but they are just one type of road information.
When lane markings are clear, they can provide very explicit lane references; when lane markings fail, the system needs to utilize road boundaries and drivable spaces to re-establish driving references. These two approaches do not replace each other but provide different levels of information for autonomous vehicles. Therefore, the disappearance of lane markings does not mean that autonomous driving has lost the road. The real test is whether the vehicle can still determine where the road is and where it can go in the absence of explicit road markings.
#AutonomousDriving #LaneMarkingRecognition