In conclusion, the number of lines is just one dimension in the performance index system of LiDAR. It is closely coupled with ranging accuracy, scanning architecture, algorithm design, electrical and thermal management, vehicle cost and even regulatory compliance. It is far from a simple matter of "more is better".
The term "line" refers to the number of discrete laser beams arranged vertically in a lidar system. Taking a mechanically rotating lidar as an example, the transmitting module fires several narrow-pulse laser beams at a fixed elevation angle, and then a rotating mechanism performs a 360° scan in the horizontal direction. Each beam then forms a spiral scanning trajectory within one revolution. A higher line count means a denser discrete angle in the elevation direction, resulting in higher vertical resolution of the point cloud and typically an expanded vertical field of view (VFOV). For MEMS or prism-oscillating solid-state lidars, although there is no mechanical rotating mechanism, multiple laser beams at fixed angles are still arranged in the elevation direction via an internal optical array, and the concept of line count still applies. Therefore, regardless of whether it's a mechanically rotating or solid-state lidar, the essence of line count is the vertical angle sampling density, which determines the magnitude of contour detail that can be captured in the vertical direction within a single frame of point cloud.
At the perception algorithm level, the increase in line count primarily leads to improved point cloud density per frame. Point cloud density directly affects the ability to reconstruct obstacle shapes, especially for targets with narrow cross-sections or at long distances; increased point cloud density means that more details can be detected. When the pitch discretization angle is reduced, pixels that were originally located in "holes," such as low curbs, fallen pedestrians, and scattered debris, will be covered by the new laser beam, allowing convolutional or Transformer networks to obtain coherent geometric information without excessive interpolation. Secondly, VFOV expansion means that within the same time window, sensors can simultaneously perceive the road surface and elevated structures, such as ramps, the bottom of overpasses, or signs suspended at intersections, thereby reducing blind spots and the risk of missed detections. For advanced autonomous driving functions such as urban NOA, Robotaxi operation platforms, and real-time relocalization scenarios using high-precision maps, the increased detail brought by a high line count not only improves obstacle detection accuracy but also reduces spatial semantic segmentation errors, thereby enhancing the robustness of the vehicle's longitudinal and lateral decision-making in complex 3D environments.
For perception systems, increasing the number of LiDAR beams can effectively improve perception performance. However, from the perspective of the entire vehicle system, increasing the beam count has far-reaching consequences. Firstly, there's the issue of hardware cost. Whether using discrete pin solutions or photonic chip integration, doubling the beam count means simultaneously expanding the capacity of the transmitting array, receiving devices, amplification, and demodulation circuits. While high-beam-count solid-state solutions offer size advantages, they require multimode miniaturization packaging and complex fiber alignment processes, resulting in more stringent yield and consistency requirements, thus significantly increasing prices during mass production. Secondly, there are the challenges of power consumption and thermal management. More laser beams mean higher pulse repetition frequencies and peak currents, leading to a doubling of heat generation and requiring larger heat sinks or active fans. For a 48V pre-installed power supply network, multiple high-beam-count LiDARs operating simultaneously may approach the fuse limit, triggering power management issues for the entire vehicle. Furthermore, data bandwidth also becomes a bottleneck. When the number of point clouds surges from over 200,000 per second to millions, gigabit Ethernet, in-vehicle PCIe bus, SoCI/O channels, and DDR bandwidth may all become congested. It is evident that handing over all point clouds to the algorithm without filtering often forces the main control computing platform to be upgraded, thereby further increasing the vehicle's BOM.
The more subtle costs occur on the algorithm side. 3D object detection and semantic segmentation often employ sparse convolution, voxel projection, or Transformer structures based on neighborhood clustering, whose time complexity increases nearly linearly. When the number of input points doubles, inference latency and memory overhead also rise accordingly. To maintain the 100ms-level closed-loop latency commonly required in mass-produced vehicles, engineers are often forced to downsample or randomly drop points at the algorithm front end. This is tantamount to using software to "offset" part of the increased number of lines in the hardware, resulting in a slowdown in the benefits.
Furthermore, a higher line count also increases the risk of electromagnetic and optical crosstalk. LiDAR typically operates in the 905nm or 1550nm band. When multiple high-line-count LiDARs of the same model in a convoy simultaneously emit pulses, the probability of time-slice conflicts between different vehicles increases. Without random time-division multiplexing, frequency-division coding, or diversified modulation schemes, the sensing results are easily mistriggered. Angular resolution is not solely determined by the line count; horizontal scanning speed, pulse divergence angle, and detection range coupling are equally important. If dense sampling is performed in the vertical direction while neglecting horizontal axis accuracy, distant targets in high-speed scenarios may still appear as sparse point clouds.
Returning to functional requirements, different levels of autonomous driving systems exhibit varying sensitivities to line count. For high-speed L2 scenarios, which prioritize lateral stability and longitudinal following, targets are primarily vehicles and guardrails with minimal height differences; 32 or 64 lines are sufficient to cover a ±10°-15° pitch field of view. In this case, further increasing the line count offers limited economic benefits, and systems engineering teams prioritize power consumption and cost. For urban L3 scenarios, dealing with traffic sign poles, pedestrians, multi-level interchanges, and low-speed dynamic targets, 128 lines or more provide a high-density point cloud that significantly reduces the false negative rate. However, if cameras and millimeter-wave radar in the fusion architecture handle some longitudinal and speed-related perception, a medium-to-high line count combined with multi-sensor fusion can balance the budget. As for Robotaxi and last-mile delivery vehicles, their business models rely on long-term operation and remote monitoring, demanding extremely high requirements for perception redundancy and safety margins; therefore, placing two to three 128-line or higher radars on the top remains reasonable within the cost framework. However, if the same approach is applied to the private car market, consumers' sensitivity to price and after-sales maintenance will make automakers rethink the rationale for "over-equipping" their vehicles.
For point cloud gradients ranging from sparse to dense, different algorithmic strategies exhibit vastly different tolerances to input size. Voxel-based sparse convolutional networks typically reach their accuracy bottleneck in low-density regions, with the gains diminishing rapidly as more point clouds are added. While full graph networks based on raw point clouds are more sensitive to density, they are often speed-limited in practical applications due to memory consumption and real-time constraints. The recently popular BEV (Bird's EyeView) projection method projects the point cloud onto a 2D plane in height layers, with a resolution threshold of approximately ten points per pixel. Beyond this threshold, the new features introduced by increasing the number of lines are likely to fall into existing pixel grids, resulting in near-zero information increment. This means that, given a computational budget, it is more cost-effective to use a limited number of lines in key regions through dynamic voxel size, adaptive sampling, and specific ROI refinement strategies than to indiscriminately increase the number of lines.
From a product lifecycle perspective, increasing the number of laser lines also increases calibration difficulty and maintenance costs. Multi-channel systems require more precise extrinsic parameter calibration. Minor pose shifts caused by mechanical wear or thermal expansion and contraction can lead to internal misalignments between multiple laser beams, necessitating more frequent in-situ self-calibration. Optical window contamination and lens coating attenuation are particularly sensitive to high-line-count radars because the transmission power of each beam is relatively low. To maintain the same signal-to-noise ratio, hardware often compensates for energy loss by increasing the pulse repetition rate, further exacerbating thermal design pressures. A high line count also means an increase in the number of laser pulses per unit time, requiring strict compliance with the IEC 60825 eye safety standard. When multiple radars operate in the same vehicle, the cumulative radiation values must also be included in the OEM's hazard assessment list. For models targeting the global market, there are also differences in laser radiation limits and application procedures between the EU, North America, and East Asia.
Because upgrading the line count results in diminishing marginal returns and soaring system costs, the industry has begun to turn to other dimensions for optimization. Wavelength upgrades are one of the mainstream approaches. The 1550nm band can output higher peak power within the eye safety threshold, thus achieving a longer detection range under the same line count conditions; at the same time, the 1550nm beam has lower atmospheric scattering attenuation, maintaining point cloud density in rain and fog environments. Another approach is to adopt a frequency modulated continuous wave (FMCW) architecture, directly measuring Doppler velocity information in a single beam, enabling medium-line-count radars to filter false alarms through joint range and velocity characteristics. Another approach is the area array Flash solution, which uses a two-dimensional SPAD array to complete the entire depth map in a single exposure and acquisition, replacing linear array rotation with pixel-level parallelism, avoiding complex mechanical and heat dissipation architectures; although the current point count is lower than that of high-line mechanical radars, it has already shown a cost advantage in the low-speed parking to L3 urban range. The common goal of all these innovations is to find the "optimal metastable state" between hardware investment, system computing power, and perception accuracy, rather than simply stacking line counts.
For a mass-produced vehicle with a planned lifecycle of three to five years, determining the appropriate number of production lines requires a closed-loop evaluation from requirements engineering to cost and factory operations. The first step is functional breakdown, clarifying the priority and safety level of target functions such as high-speed cruise control, city NOA (Noise, Arrival, and Automation), automatic parking, and even short-term L4 autonomous driving. The second step is to assess the vehicle's computing power and electronic/electrical architecture, determining whether the existing ECU bandwidth and power reserves can handle the transmission and inference of higher-density point clouds. The third step considers the vehicle's mounting location and styling constraints; a top-mounted design allows for 360° surround view but significantly impacts wind resistance and overall vehicle height; fender-embedded or front grille-hidden designs require a wider VFOV (Visible Field of View) to offset obstruction. The fourth step involves supply chain and mass production feasibility analysis; high-line-count solid-state solutions are often in the ramp-up phase, and delivery yield, spare parts costs, and after-sales service capabilities all need to be verified. Finally, using the Total Cost of Ownership (TCO) as a benchmark, hardware prices, software development, calibration and debugging, long-term maintenance, and cloud data storage are comprehensively calculated to ensure that the selected number of production lines has a positive profit margin in the business model.
Line count is undoubtedly the most intuitive and easily amplified parameter for LiDAR in marketing. It does indeed provide richer vertical information, helping algorithms capture minute targets and complex structures. However, as autonomous driving increasingly moves towards mass production, "point cloud accuracy" must be considered within a vast system trade-off matrix: hardware cost, power and thermal design, data links, algorithm latency, regulatory compliance, maintenance and upkeep, and even consumer willingness to pay—no factor can be ignored. Beyond a certain threshold, the marginal benefit of line count rapidly declines, potentially triggering a chain reaction of problems such as power overload, computing bottlenecks, and regulatory risks.
Instead of blindly piling on components, resources should be invested in higher power, longer wavelengths, and more intelligent anti-interference modulation, as well as adaptive sampling and deep fusion of multiple sensors on the algorithm side. Only when the number of hardware lines, system computing power, and software algorithms reach a dynamic balance can LiDAR truly become the "third eye" of autonomous driving rather than an expensive burden; and only in such a balance can every laser beam be used to its fullest potential, allowing vehicles to operate more safely and efficiently on real roads.
