Light detection and ranging (LiDAR) is a competitive technique for acquiring three-dimensional (3D) information, which has been widely used in fields like terrain mapping^[1,2], robotics^[3,4], target detection^[5,6], and autonomous driving^[7–9]. Time-of-flight LiDAR, with its simple structure, low cost, and long working distance^[10], is the preferred choice for autonomous driving. Flash LiDAR, illuminating the entire field of view (FOV) and employing a detector array for detecting, can rapidly generate high-resolution point clouds^[11,12]. However, it faces challenges such as limited FOV and working distance, high cost, and non-uniform pixels^[13]. Scanning LiDAR acquires 3D information within the FOV by steering the laser direction and is more suitable for long-range imaging. It can be categorized into two major categories: solid-state and mechanical types. Solid-state scanning devices, represented by optical phased arrays (OPAs), typically offer extremely high scanning speeds. However, they are not yet mature, usually featuring limited scanning ranges and high costs^[14]. In contrast, mechanical scanning, though slower, is robust and well-developed, making it the predominant choice for current automotive LiDAR^[14–17]. For example, the Falcon by Seyond employed a rotating mirror and a galvanometer mirror as the fast and slow axes of scanning, respectively^[18–20]. With the multi-beam technique in use, the LiDAR could scan multiple laser beams simultaneously, which increased the acquisition rate of spatial information^[21,22]. Li et al. achieved 3D imaging with a pixel of 80×1200 with 100-beam scanning^[23]. In the current automotive LiDAR products, the multi-beam technique was applied to supplement the scanning in the direction of laser source arrangement, improving the LiDAR resolution or even replacing the slow-axis scanning completely, such as the Velodyne VLS-128, Hesai AT128, and Robosense M series^[24–26]. Table 1 lists the main parameters of the above products and some related research^[27,28]. The worse the resolution, the higher the possibility for the LiDAR to miss small targets, thereby compromising safety. In the listed LiDAR systems, the arrangement of sampling points typically exhibited a grid-like pattern, as illustrated in Fig. 1(a). Notably, these sampling points do not fully encompass the FOV. By augmenting the repetition frequency, the potential points depicted in Fig. 1(a) could enhance the resolution along the fast-axis direction. However, the line spacing within the slow-axis direction remained constant. The line spacing in linear scanning primarily depends on factors such as the scanning velocity of the slow axis, the quantity of laser beams, and the extent of the FOV.

The predominant strategy to diminish line spacing and enhance vertical resolution involved augmenting either the number of laser beams or the scanning speed. However, both approaches necessitated meeting the performance constraints of the hardware, thereby posing certain requirements. Conducting more refined scanning in the region of interest (ROI) could bring smaller sampling point spacing. However, this approach involved a compromise between FOV and spatial resolution. Specific solid-state scanners, notably OPA, demonstrate remarkably rapid scanning capabilities. Yet, they still face challenges in terms of cost, stability, and so on, impeding their independent deployment as scanning devices^[29,30].

In this work, we proposed a new scanning method for automotive scanning LiDAR. Based on the traditional linear scanning method, an acousto-optic modulator (AOM) was applied to supplement small-angle, high-frequency scanning, which enhanced the LiDAR’s ability to distribute sampling points and eliminated the impact of line spacing on vertical resolution. When the laser pulse repetition frequency is high and considering the beam size, it possesses the capability to eliminate all blind spots within the FOV, as illustrated in Fig. 1(b). Consequently, the resolution is solely dependent on the divergence angle of the laser beam. We conducted experiments to verify the actual effect of this scheme on improving the LiDAR’s ability to recognize small targets. The results show that this scanning scheme increased the probability of recognizing thin tubes and flat boards at 40 m by more than 20% to over 90% with the spatial resolution of 0.03° (0.5 mrad).

AOM is an optical device that uses the interaction between sound waves and light waves to modulate the intensity, phase, or polarization of light. The diffraction angle is proportional to the frequency of the wave applied to the acousto-optic crystal. By adjusting the frequency, the direction of the beam can be steered^[31,32]. Despite its exceptionally high scanning speed, the AOM typically has a limited scanning range and aperture, rendering it unsuitable for standalone use as a scanner in LiDAR systems. In the proposed method, the AOM was used to supplement a high-frequency scanning at the mrad level in the slow-axis direction, thereby significantly enhancing the LiDAR’s capability to distribute sampling points and improving its resolution.

Figure 2 shows the schematic diagram of the experimental setup. The system operated in the near-infrared 1550 nm wavelength and consisted of a quasi-coaxial optical structure. For the laser emission optical path, a laser diode (LD) with a central wavelength of 1550 nm outputs a pulsed laser seed with a repetition frequency of 3 MHz and a pulse width of 1.4 ns. After amplification by an erbium-doped fiber amplifier (EDFA) and collimation by a collimator, the beam waist diameter was approximately 0.8 mm, with an output laser power of 40 mW. A half-wave plate (HWP) was used to adjust the polarization direction of the laser, ensuring that the majority of the energy could pass through the polarization beam splitter (PBS). Then, the beam underwent a small-angle scanning via an AOM. The AOM was driven by a frequency modulation (FM) signal, whose instantaneous frequency varied periodically with time to achieve rapid scanning. After a 5× beam expander (Thorlabs GBE05-C), the diffracted laser from the AOM exhibited a beam waist diameter of about 4 mm with the divergence angle decreased to 0.5 mrad. Additionally, the scanning angle of the laser direction modulated by the AOM was also diminished to 1/5. A PBS and two 2D scanning galvanometers (G1, G2) were utilized to achieve the traditional linear scanning. A small part of the laser energy was reflected by the PBS and detected by a photodetector, and the intensity variation could be used as a synchronization signal for the AOM scanning. The receiving aperture of the galvanometer mirrors was 10 mm. Regarding the echo photon receiving path, the emitted laser illuminated the target with a small fraction of the diffuse reflection echo returning to the PBS along the original path. A part of the echo was reflected by the PBS and coupled into a multimode fiber with a core diameter of 62.5 µm, which was connected to the single-photon detector. The lens had a focal length of 25 mm, and the receiving FOV was 2.5 mrad. The detector was a homemade GHz-gated InGaAs/InP single-photon detector, exhibiting a photon detection efficiency of 3% and a dead time of 16 ns.

The AOM (MGAS110-A1, Opto-Electronic) utilized in the experiment had a rated radio frequency (RF) drive of 110 MHz. When the RF frequency ranged between 104 and 115 MHz, the diffraction efficiency was no less than 30%. In the experiment, a triangular wave with a frequency of 200 kHz was employed to modulate the frequency of a sine wave carrier with a frequency of 110 MHz to drive the AOM. The frequency deviation D was set at 6 MHz. The instantaneous frequency f(t) and the expression of the FM signal g(t) are as follows: f(t)=fc+(−1)⌊2tTm⌋·(4tTm−2·⌊2tTm⌋−1)·D,(1)g(t)=Asin[∫0tf(t)dt+θ0],(2)where fc is the center value of the instantaneous frequency (110 MHz), Tm is the period of the modulating signal (the period of the instantaneous frequency variation), A is the amplitude of the FM signal, and θ0 is the initial phase of the FM signal. Figure 3(a) shows the waveform of the FM signal and the schematic diagram of the instantaneous frequency varying with time. With this FM signal, the diffracted laser was scanned at a frequency of 200 kHz. Following the beam expander (BE), the actual scanning angle of the outgoing laser, modulated by the AOM, was measured to be 1.48 mrad. The sum of this scanning angle and the divergence angle of the outgoing laser was within the receiving FOV. The spatial relationship among the three is shown in Fig. 3(b). This numerical relationship made the quasi-coaxial optical path feasible. A sine wave operating at a frequency of 160 Hz drove the fast-axis galvanometer with an optical angle of 10°. A 32-step wave drove the slow-axis galvanometer with a frequency of 5 Hz, resulting in a frame rate of 10 frames per second (fps). Each step of the galvanometer had an optical angle of 0.09°, corresponding to the line spacing. This performance is comparable to that of high-performance automotive LiDARs that are already in mass production. Although the FOV in the experiment was smaller than that of the automotive LiDAR products, the scanning principles were the same. The distinction resided in the scanning devices and the number of lasers. Therefore, this method is universally applicable to automotive LiDARs.

During data processing, this novel method does not introduce any additional steps. By integrating a term associated with AOM scanning into the standard linear scanning process of automotive LiDAR, the spherical coordinates of echo points can be accurately calculated, {d=c(t−tp)2α=4Ag·fg|t−tg−12fg|−Agβ=4Am·fm|t−tm−12fm|−Ac+⌊2fgt⌋·θstep.(3)

In Eq. (3), c is the speed of light, t is the moment when the echo was detected, tp, tg, and tm represent the moments of the most recent synchronization signals of the laser pulse, fast axis galvanometer, and AOM, respectively. Ag, Am, fg, and fm are the amplitudes and frequencies of the fast-axis galvanometer and the AOM scanning, respectively. θstep is the step angle of the slow-axis galvanometer. d, α, and β correspond to the distance, elevation, and azimuth angles in spherical coordinates. Then, the spatial position of the echo points can be obtained through coordinate transformation. The main error in the imaging system when reconstructing the 3D coordinates of the echo signal lies along the laser propagation direction, which is caused by the width of the laser pulse, {x=d·cosβsinαy=d·sinβz=d·cosβcosα.(4)

To demonstrate the effectiveness of this method, we first compared the recognition probability of small targets in indoor settings, utilizing single-line LiDAR with AOM supplementary scanning against traditional linear scanning. As shown in Figs. 4(a) and 4(b), during the experiment, these tubes placed on a table were utilized as targets, with their positions extending beyond the edge of the table, simulating vehicles carrying reinforcing steel bars in potential autonomous driving scenarios.

In Fig. 4(a), the tubes were laid flat and represented Target 1, while Target 2 in Fig. 4(b) had a 10° elevation angle. Each tube has a diameter of 2.8 cm, so four pipes placed side by side correspond to an angle of only 0.16°×0.04° in the FOV at 40 m. The black marking on the board, labeled as P in Figs. 4(a) and 4(b), appeared at nine different positions in the FOV, as shown in Fig. 4(c). The bias of the slow-axis galvanometer was chosen with two slightly different values, which is equivalent to 18 different positions within the FOV. A total of 25 frames of point cloud were obtained using two scanning methods at each position, and the probability of recognizing the target was statistically analyzed.

The results are shown in Fig. 5(a). AOM supplementary scanning significantly enhanced the LiDAR’s ability to recognize the board and tubes, especially when the targets were located at column 2 and column 3, the recognition probability generally increased by more than 20%, reaching over 90%. The recognition probability did not reach 100%, as shown in the scanning trajectory diagram, because of the insufficient echo points. When the instantaneous frequency of the FM signal was far from 110 MHz, the AOM’s diffraction efficiency dropped significantly. Moreover, the fast-axis galvanometer was driven by a sine wave, so the sampling point density at column 1 was lower than that of the other two columns. This led to the possibility that, even if the scanning trajectory covered the target, the target might not be successfully detected due to an insufficient number of echo points. The trend of recognition probability changes in different columns also proves this point. When using traditional LiDAR linear scanning, the recognition probability was almost independent of the sampling point density, as blind areas in the scanning trajectory did not decrease with increasing repetition frequency. However, when using AOM supplementary scanning, the recognition probability increased with the increase of sampling point density. It can be anticipated that, with sufficient echo points, this new scanning method will perform better, and the recognition probability of the target will approach 100%.

There was no significant difference between the recognition probability of target 1 and target 2. Although the tubes in Fig. 4(b) had a larger cross-sectional area, the echo points generated by the inclined plane at a small angle to the laser emission direction had a weak spatial correlation. For example, with an angle of 10°, it is equivalent to reducing the point density in the point cloud image by cot10°≈5.7 times, making it easy to identify these echo points as noise and remove them during the denoising process. This further highlights the safety hazards posed by such targets in practical applications and the importance of improving resolution.

In addition to the above scenarios, we also conducted a set of imaging experiments to continuously image a moving pedestrian with an umbrella at a distance of 40 m. The frame rate in these two experiments was reduced to 2 fps to achieve higher sampling point density and highlight the differences between scanning methods. Comparing one frame from the point cloud videos, as shown in Fig. 6, it can be found that the imaging results using AOM supplementary scanning better restored the details of the targets, such as the umbrella in the pedestrian’s hand. This intuitively proves that this scanning method improves the resolution of the LiDAR. It is worth noting that, due to the continuous motion of pedestrians, the postures of the pedestrian in the two point clouds are not identical, while the umbrella maintains a consistently horizontal orientation. Traditional linear scanning LiDAR failed to detect the umbrella held by the pedestrian in many frames, posing potential safety hazards.

We then imaged an intersection outdoors to replicate the real-life application scenario of automotive LiDAR. A tricycle was parked on the side of the intersection with some tubes placed on it, extending beyond the rear of the tricycle. Figures 7(a) and 7(b) are the photos of the actual scene. The parameters of the experimental setup are the same as the previous experiment, except that the frame rate is reduced by half to 5 fps and the horizontal FOV is doubled to 20°, which maintains the same sampling point density. By analyzing the imaging results of two methods, totaling 40 frames of point clouds, it was found that compared to traditional linear scanning, the AOM supplementary scanning increased the probability of recognizing tubes from 85% to 100%. The recognition probability of the thin board laid flat on the vehicle was also improved. Figures 7(c)–7(f) show one frame from each of the two sets of imaging results. From the partial projection images in Fig. 7(f), it can be intuitively seen that the traditional automotive LiDAR linear scanning method still has a probability of completely missing targets because of the sampling points distribution in blind areas with a line spacing of 0.09°. The new scanning method brought a noticeable improvement in resolution and target details, consistent with the results of the indoor simulation experiments. Moreover, the point cloud obtained by AOM supplementary scanning has a better denoising performance under the same K-nearest neighbor (KNN) denoising parameters. This is also due to the more uniform distribution of sampling points.

This work proposes a new scanning scheme for LiDAR in the application scenarios of autonomous driving. The scanning scheme uses an AOM to add high-speed small-angle scanning in the slow-axis direction of traditional linear scanning, significantly improving the distribution of sampling points. It eliminates the blind areas in the FOV, resulting in a resolution of 0.03°, equal to the laser divergence angle, and increases the recognition probability of small flat targets such as flatbeds, horizontal rods, and cargo extending beyond the rear of vehicles. Compared with the currently available high-performance automotive LiDAR products, this method increases the recognition probability of flat targets no higher than 3 cm at 40 m (less than 0.04°) from 70% to over 90%. When a vehicle travels at a speed of 80 km/h, detecting the target 0.5 s earlier can advance the braking position by over 11 m. This work demonstrates the effectiveness and feasibility of using a high-speed, small-aperture, and small-angle scanner in conjunction with a slow-speed, large-aperture, and large-angle scanner to achieve high-resolution imaging. The galvanometers in the experiments in this paper can be replaced with other scanners, like rotating mirrors and micro electro mechanical system (MEMS) mirrors, to meet the requirements for product commercialization. There is also room for further optimization of AOMs in future work, such as using AOMs with higher diffraction efficiency or other solid-state scanners with higher integration and lower cost. The AOM has no moving parts and consumes relatively low power, which endows it with excellent thermal stability and shock resistance. Its compact size also allows it to be easily integrated into small LiDAR products. This scheme provides a feasible method for achieving no blind areas in automotive LiDAR, providing an effective approach for the application of LiDAR in more advanced autonomous driving.