# An Asynchronous Peak Finding Approach for Neuromorphic Depth Sensing in Flash LiDAR

## Abstract

In this paper, we present a novel asynchronous peak detection method for Direct Time-of-Flight (dToF) flash LiDAR systems. The approach allows to adjust the exposure time in each pixel based on thresholding the maximum of the histogram with a statistical confidence level. By asynchronously reporting peak events instead of entire frames, our approach significantly reduces latency and enhances photon efficiency. Implemented on a Xilinx Kintex-7 FPGA and tested on a 128x64 SPAD array, the method demonstrates improved motion blur mitigation. This technique holds promise for applications such as autonomous vehicles and augmented reality, where rapid and accurate depth sensing is essential.

## Authors

Yiyang Liu *Institute for Integrated Micro and Nano Systems, School of Engineering, The University of Edinburgh, Edinburgh, UK*

Sarrah M. Patanwala *Institute for Integrated Micro and Nano Systems, School of Engineering, The University of Edinburgh, Edinburgh, UK*

Alistair Gorman *Institute for Integrated Micro and Nano Systems, School of Engineering, The University of Edinburgh, Edinburgh, UK*

Istvan Gyongy *Institute for Integrated Micro and Nano Systems, School of Engineering, The University of Edinburgh, Edinburgh, UK*

Robert K. Henderson *Institute for Integrated Micro and Nano Systems, School of Engineering, The University of Edinburgh, Edinburgh, UK*

## Publication Information

**Journal:** 2025 IEEE International Symposium on Circuits and Systems (ISCAS) **Year:** 2025 **Pages:** 1-5 **DOI:** [10.1109/ISCAS56072.2025.11043772](https://doi.org/10.1109/ISCAS56072.2025.11043772) **Article Number:** 11043772 **ISSN:** Electronic ISSN: 2158-1525, Print on Demand(PoD) ISSN: 0271-4302

## Metrics

**Total Downloads:** 111

---

## Keywords

**IEEE Keywords:** Laser radar, Three-dimensional displays, Neuromorphics, Sensors, System-on-chip, Vehicle dynamics, Autonomous vehicles, Augmented reality, Photonics, Single-photon avalanche diodes

**Index Terms:** Depth Camera, Flash Lidar, Autonomous Vehicles, Peak Detection, Motion Blur, Lidar System, Single-photon Avalanche Diode, Pixel Time, Square Root, Background Levels, Photon Counting, Depth Measurements, Central Limit Theorem, Lidar Data, Histogram Data, Histogram Bins, Bin Values, Object Depth, Background Peak, Dynamic Vision Sensor, Time-to-digital Converter

**Author Keywords:** 3D ranging, light detection and ranging(LiDAR), single photon avalanche diode(SPAD), neuromorphic sensing

undefined
## SECTION I. Introduction

Direct Time-of-Flight (dToF) LiDAR systems have become a popular choice for depth-sensing in applications like autonomous driving, auto-focusing, and Augmented Reality (AR) [^1]. They provide precise distance measurements with millimeter resolution, making them essential in these fields. However, the conventional LiDAR relies on mechanical components for scanning, which increases complexity and introduces motion artifacts. This in turn can affect the accuracy of depth measurements in dynamic environments. A promising alternative is flash LiDAR, which uses a pulsed laser to illuminate the entire scene and an array-based photon-counting receiver to capture photons from all pixels simultaneously [^2] [^3]. This design eliminates mechanical moving parts and time differences between pixels, reducing motion artifacts and improving accuracy.

The exposure time for conventional synchronous flash LiDAR is set to meet the requirement of detecting objects under nominal worst-case conditions. However, the varying depth and reflectivity of objects across the different pixels can cause the speed of histogram formation to differ within the array. The pixels with objects closer to the LiDAR or with higher reflectivity may already have a complete depth histogram but must wait for the other pixels until the end of the exposure period. This introduces extra latency and increases the risk of histogram corruption if the target moves during this time. Therefore, asynchronously reporting depth information as soon as the peak is identified in the histogram data, could help address the latency issue and improve system reliability in dynamic environments.

In recent years, neuromorphic vision sensors like Time-to-First-Spike (TTFS) sensors [^4] and Dynamic Vision Sensors (DVS) [^5] have gained attention for their high dynamic range, low latency, and power efficiency. These sensors typically include an event generation circuit in each pixel and an Address Event Representation (AER) module for pixel readout. Similar concepts have been adapted to Single Photon Avalanche Diode (SPAD) image sensors. In [^6], a dynamic vision event generator was implemented in SPAD pixels using a logarithm-2-based thresholding module, and in [^7], dynamic vision event generation was applied through post-processing the 4-bit raw image. However, these implementations focused on SPAD-based neuromorphic vision sensing without utilizing the SPAD’s timing and depth detection capabilities. In [^8], the proposed flash LiDAR receiver reports depth changes detected by pixels, and [^9] presents a SPAD-based TTFS sensor with post-processed DVS-type event generation and 3D depth measurement. However, the former work only generates dynamic depth data in frames, and the latter performs depth measurement through time gates which is both time and photon inefficient.

In this paper, we introduce a novel neuromorphic depth sensing approach for dToF flash LiDAR systems. To the best of our knowledge, this is the first application of neuromorphic principles to a dToF LiDAR. Our method enables asynchronous peak detection by continuously monitoring the histograms within each pixel, and generating the peak detection events based on thresholding the maximum of the histogram with a statistical confidence level, thus allowing the LiDAR data to be readout asynchronously. We implemented this approach on a Xilinx Kintex-7 Field Programmable Gated Array (FPGA), interfaced with a SPAD sensor fabricated in 40nm CMOS technology [^10].

## SECTION II. Asynchronous Peak Detection

### A. Peak Detection

In dToF LiDAR systems, a laser typically emits a pulsed beam, which is then reflected by objects in the scene. The photons from the reflected laser beam and the background arriving on the sensor are timestamped and stored in a histogram. Generally, the ambient photons are uniformly distributed across the entire time range of the histogram. Therefore, the histogram bins corresponding to ambient light, *bamb*, are expected to have the same value during a single laser cycle, *βB*, where *β* is an attenuation factor related to the Photon Detection Efficiency (PDE) of the SPAD and environmental factors such as the depth and reflectivity of an object. *B* represents the ambient light intensity per bin. However, due to photon shot noise, the values of the ambient bins are subject to Poisson fluctuations, so *bamb* ~ *Po*(*βB*).

Typically, to distinguish the reflected signal from the background light, the LiDAR system fires the laser *N* times, with *N* usually ranging from several hundred to several thousand. As a result, after *N* exposures, the value of the ambient bins becomes *bamb* ~ *Po*(*βBN*), according to the additive property of Poisson distributions. Since *N* is large, by the Central Limit Theorem (CLT), *bamb* can be approximated by a Gaussian distribution, where *bamb* ~ *N*(*βBN,βBN*). Therefore, the histogram bins corresponding to the laser return signal (histogram peaks) should exhibit a photon count as shown in (1), with a thresholding factor *α.*

$$
\begin{equation*} {b_{sig}} > {\text{ }}\beta BN + \alpha \sqrt {\beta BN} \tag{1}\end{equation*}
$$

### B. Background Estimation

![Figure 1](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu1-p5-liu-large.gif)

*Fig. 1: Example of Background Estimation.*

To achieve the peak detection described in the previous section, the background level must be estimated from the histogram data. This can be accomplished by calculating the mean or median of the background bins, or by accumulating a histogram based solely on ambient photons in the cycles where the laser is not fired. However, the former method requires additional adder and divider blocks, while the latter is time-consuming, leading to delays in data generation. In this work, we estimate the background level using a comparison-based method, which simplifies the circuit into comparators and reduces the complexity.

An example of the estimation process is shown in Fig. 1. It illustrates a LiDAR histogram with 128 bins and a ground truth background level of 32 photons per bin, indicated by the red line. If the thresholding factor *α* is set to 3, the peak detection threshold becomes $32 + 3 \times \sqrt {32} = 48.97$, represented by the red dotted line. In our proposed approach, the 128-bin histogram is divided into four quarters (Q1-Q4). The value of bins in each quarter is compared to identify the maximum of each quarter, shown as red dots in Fig. 1. The larger value from the first half (Q1 and Q2) is compared to the second half (Q3 and Q4) to locate the histogram’s peak (Q3 in this case), and define the peak (Q3-Q4) and non-peak halves (Q1-Q2). The smaller value from the non-peak half (Q2 in this example) is then used as the estimated background level, shown as the green line, with the corresponding threshold indicated by the green dotted line. In this example, the value of the peak bin exceeds the threshold, thus triggering a peak detection event. Overall, the background level estimated through this process is usually higher than the actual background level, but this is acceptable since it increases the threshold of the peak detection, and could further prevent the generation of false-positive events.

### C. Asynchronous Operation

The complete approach is shown in Fig. 2. In this operation, three parameters need to be manually decided: the threshold level parameter *α*, and the two laser cycle limits *L_lim*1 and *L_lim*2. According to (1), *alpha* is related to the level of rejection for a Normal distribution, which adjusts the rate of false-positive detections. However, increasing the value of *α* will also increase the time required to detect a potential peak, hence introducing extra latency. *L_lim*1 is a lower limit for exposure to satisfy the CLT as stated in Section II.A, which is adjusted to an appropriate number when most of the background bins contain at least 10 counts according to the background level. And *L_lim*2 is the maximum exposure cycle, which means there is no object in the scene present for this pixel if there is still no peak identified until *L_lim*2. After setting these parameters, the histogram inside each pixel is checked after each *N* laser cycle to see whether the current histogram peak has sufficient photon count to be identified as a true peak.

![Figure 2](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu2-p5-liu-large.gif)

*Fig. 2: Flowchart for the Asynchronous Operation.*

## SECTION III. Hardware Implementation

![Figure 3](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu3-p5-liu-large.gif)

*Fig. 3: Hardware Architecture.*

The proposed approach is implemented on a Xilinx Kintex7 FPGA, connected to a 128x64 SPAD sensor [^10], the system architecture is shown in Fig. 3. Due to limited I/O resources on the FPGA, the sensor was designed to report only a subsection of the array - 128 SPADs, at the same time. A complete image is then acquired by electronically scanning different regions of the array through a control signal generated from FPGA. The FPGA firmware comprises two parts: the time-of-flight part, and the processing part. The ToF module contains a Time to Digital Converter (TDC) with 250 ps bin resolution, which samples the input signal from 128 SPADs as shown in Fig. 3. The sampled signals are then multiplexed and accumulated into four 128-bin, 10-bit histograms. Note that 128 inputs are divided into 2 groups of 64, then multiplexed into four histograms each due to limited FPGA logic resources, thus requiring an additional 16× scanning to form the images shown in the experimental results.

![Figure 4](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu4-p5-liu-large.gif)

*Fig. 4: Pipelined Processing.*

The processing module is used to achieve the asynchronous background estimation and peak detection, described in Section II. Two such modules are implemented on the FPGA after the ToF module, each of which takes the histograms of 4 pixels as their input. The processing of a single histogram takes 4 laser cycles. In order to increase the time efficiency, the processing is implemented at the same time as the ToF histogram accumulation, as shown in Fig. 4. The implementation of the algorithm is composed of two stages. Stage 1 compares the 128 histogram bins for peak identification and background level estimation. The second stage computes the threshold and compares the peak count to the threshold. At the start of each processing cycle, the histogram of the selected pixel is sampled into stage 1. It is then time-multiplexed through a 16-to-1 comparator tree to observe the maxima from the consecutive groups of 16 histogram bins. At the end of the fourth laser cycle, while the system waits for the next laser trigger to be generated, the above computed 8 maxima are passed through the same comparator tree to achieve the peak and background identification described in Section II.B.

The calculated background photon count is passed to stage 2 in the next processing cycle. Here, the integral square root of the background is generated using a bitwise value estimation approach, as shown in Alg. 1. This algorithm maintains the error below 1 for all inputs, while it takes 5 clock cycles to estimate the background’s square root, as shown in Fig. 5. The threshold is then calculated using (1). Instead of implementing a multiplier, the threshold level is adjusted by setting *α* equal to either 1,2,4, or 8. This is equivalent to left-shifting the calculated square root by 1,2,3, or 4 bits.

![Figure 5](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu.al1-p5-liu-large.gif)

*Algorithm 1*

![Figure 6](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu5-p5-liu-large.gif)

*Fig. 5: Error of the Bitwise Integral Square Root Algorithm Over Different Background Levels.*

Finally, when a peak event is generated, a request (Req) signal is sent to the AER module. It responds back with an acknowledgment (Ack) signal, while simultaneously passing a write enable signal to the First In First Out (FIFO) module. The FIFO is then readout to the host PC for further data post-processing.

![Figure 7](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu6-p5-liu-large.gif)

*Fig. 6: LiDAR System.*

## SECTION IV. Experimentation

To validate the proposed approach, a flash LiDAR system using the FPGA, the sensor, and a 773nm Hamamtsu M1030615 laser head (with pulse width as 120ps and 10MHz repetition rate), is assembled as shown in Fig. 6. This setup is used to image a fan controlled by a stepper motor, as shown in Fig. 7a. Fig. 7b shows the captured intensity data from the sensor. The fan is synchronized with the LiDAR system to maintain the same motion for each electronic scan. This helps mimic the performance of the proposed approach as in typical flash LiDAR systems.

![Figure 8](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu7-p5-liu-large.gif)

*Fig. 7: Scene for Imaging, Captured by: (a) RGB Camera, (b) Sensor in Intensity mode.*

### A. Static Fan Measurement

Firstly, the LiDAR data is captured while the fan is static, and the results are presented in Fig. 8. In this experiment, the parameter *α* was chosen as 8, *L_lim*1 was set to 1300 laser cycles, and *L_Lim*2 was set to 10000 laser cycles. In Fig. 8a, each point represents a peak detection event generated from the system. It is plotted in a spatial and temporal plot where the color shows the depth of the object corresponding to the peak detected from the histogram bins. The four slices shown in Fig. 8b,c,d,e are the detected peak events, summed at laser cycles 0, 0 to 1500, 0 to 3000, and 0 to 6000 respectively. This demonstrates the asynchronous formation of the depth maps. It is observed that the number of pixels detecting a peak increases over time. Additionally, the pixels corresponding to the fan detect and report peaks more frequently than those imaging the background.

![Figure 9](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu8-p5-liu-large.gif)

*Fig. 8: Peak Detection Event Plot of the Static Fan.*

### B. Dynamic Fan Measurement

The second measurement is taken while the stepper motor controlling the fan is rotating at 120 rounds per minute, this is to emulate a lateral motion in the scene. All the parameters are maintained the same as in the previous experiment. The results are presented in Fig. 9. The fan is rotated clockwise. Hence in Fig. 9a, the right edge of the fan moves downwards, and the top edge moves towards the right with respect to the temporal axis. In this figure, the four slices are taken from laser cycles 5000 to 40000, each containing events generated in 5000 cycles. The motion of the fan is shown clearly across these four figures, without any motion blur. Additionally, it is worth noting that the noise in the image is from the sensor readout (fixed pattern noise), and insufficient laser power (50mW peak power).

![Figure 10](https://ieeexplore.ieee.org/mediastore/IEEE/content/media/11043142/11042930/11043772/liu9-p5-liu-large.gif)

*Fig. 9: Peak Detection Event Plot of the Rotating Fan.*

## SECTION V. Conclusion

In this paper, we presented an asynchronous peak detection method for dToF flash LiDAR systems. The continuous background monitoring and peak thresholding approach optimizes the utilization of photon counts. It allows the system to focus on significant peak events rather than continuously processing entire frames. Thus, achieving a novel neuromorphic depth sensing technique. This event-driven paradigm significantly reduces latency and enhances the effective data rate, providing a unique advantage for real-time depth sensing applications. The FPGA-based prototype successfully demonstrates an improved depth sensing capability with reduced motion blur. The proposed implementation can be integrated into on-chip SPAD sensors to facilitate efficient event-driven LiDAR data processing, making it promising for various dynamic applications such as autonomous vehicles and augmented reality.

## Footnotes

. This work is funded by EPSRC EP/S026428/1

## References

[^1]: O. Kumagai, J. Ohmachi, M. Matsumura, S. Yagi, K. Tayu, K. Amagawa, T. Matsukawa, O. Ozawa, D. Hirono, Y. Shinozuka, R. Homma, K. Mahara, T. Ohyama, Y. Morita, S. Shimada, T. Ueno, A. Matsumoto, Y. Otake, T. Wakano, and T. Izawa, “7.3 A 189×600 Back-Illuminated Stacked SPAD Direct Time-of-Flight Depth Sensor for Automotive LiDAR Systems,” in 2021 IEEE International Solid-State Circuits Conference (ISSCC), vol. 64, Feb. 2021, pp. 110–112. [IEEE](https://ieeexplore.ieee.org/document/9365961) [Google Scholar](https://scholar.google.com/scholar?as_q=7.3+A+189%C3%97600+Back-Illuminated+Stacked+SPAD+Direct+Time-of-Flight+Depth+Sensor+for+Automotive+LiDAR+Systems&as_occt=title&hl=en&as_sdt=0%2C31)

[^2]: R. K. Henderson, N. Johnston, S. W. Hutchings, I. Gyongy, T. A. Abbas, N. Dutton, M. Tyler, S. Chan, and J. Leach, “5.7 A 256×256 40nm/90nm CMOS 3D-Stacked 120dB Dynamic-Range Reconfigurable Time-Resolved SPAD Imager,” in 2019 IEEE International Solid-State Circuits Conference - (ISSCC), Feb. 2019, pp. 106–108. [IEEE](https://ieeexplore.ieee.org/document/8662355) [Google Scholar](https://scholar.google.com/scholar?as_q=5.7+A+256%C3%97256+40nm%2F90nm+CMOS+3D-Stacked+120dB+Dynamic-Range+Reconfigurable+Time-Resolved+SPAD+Imager&as_occt=title&hl=en&as_sdt=0%2C31)

[^3]: D. Stoppa, S. Abovyan, D. Furrer, R. Gancarz, T. Jessenig, R. Kappel, M. Lueger, C. Mautner, I. Mills, D. Perenzoni, G. Roehrer, and P.-Y. Taloud, “A Reconfigurable QVGA/Q3VGA Direct Time-of-Flight 3D Imaging System with On-chip Depth-map Computation in 45/40nm 3D-stacked BSI SPAD CMOS,” in 2021 International Image Sensor Workshop (IISW). [Google Scholar](https://scholar.google.com/scholar?as_q=A+Reconfigurable+QVGA%2FQ3VGA+Direct+Time-of-Flight+3D+Imaging+System+with+On-chip+Depth-map+Computation+in+45%2F40nm+3D-stacked+BSI+SPAD+CMOS&as_occt=title&hl=en&as_sdt=0%2C31)

[^4]: S. Chen, A. Bermak, and D. Martinez, “A CMOS vision sensor with on-the-fly histogram equalization using TFS encoding and AER readout,” in 2005 IEEE Asian Solid-State Circuits Conference, Nov. 2005, pp. 433–436. [IEEE](https://ieeexplore.ieee.org/document/4017624) [Google Scholar](https://scholar.google.com/scholar?as_q=A+CMOS+vision+sensor+with+on-the-fly+histogram+equalization+using+TFS+encoding+and+AER+readout&as_occt=title&hl=en&as_sdt=0%2C31)

[^5]: P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128\times 128 120 dB 15 µ s Latency Asynchronous Temporal Contrast Vision Sensor,” IEEE Journal of Solid-State Circuits, vol. 43, no. 2, pp. 566–576, Feb. 2008. [IEEE](https://ieeexplore.ieee.org/document/4444573) [Google Scholar](https://scholar.google.com/scholar?as_q=A+128%5Ctimes+128+120+dB+15&as_occt=title&hl=en&as_sdt=0%2C31)

[^6]: F. Mattioli Della Rocca, H. Mai, S. W. Hutchings, T. A. Abbas, K. Buckbee, A. Tsiamis, P. Lomax, I. Gyongy, N. A. W. Dutton, and R. K. Henderson, “A 128 × 128 SPAD Motion-Triggered Time-of-Flight Image Sensor With In-Pixel Histogram and Column-Parallel Vision Processor,” IEEE Journal of Solid-State Circuits, vol. 55, no. 7, pp. 1762–1775, Jul. 2020. [IEEE](https://ieeexplore.ieee.org/document/9097202) [Google Scholar](https://scholar.google.com/scholar?as_q=A+128+%C3%97+128+SPAD+Motion-Triggered+Time-of-Flight+Image+Sensor+With+In-Pixel+Histogram+and+Column-Parallel+Vision+Processor&as_occt=title&hl=en&as_sdt=0%2C31)

[^7]: K. Morimoto, N. Isoda, H. Sekine, T. Sasago, Y. Maehashi, S. Mikajiri, K. Tojima, M. Shinohara, A. Abdelghafar, H. Tsuchiya, K. Inoue, S. Omodani, K. Chida, A. Ehara, J. Iwata, T. Itano, Y. Matsuno, K. Sakurai, and T. Ichikawa, “3D-Stacked 1Megapixel Time-Gated SPAD Image Sensor with 2D Interactive Gating Network for Image Alignment-Free Sensor Fusion,” in 2024 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits), Jun. 2024, pp. 1–2. [IEEE](https://ieeexplore.ieee.org/document/10631424) [Google Scholar](https://scholar.google.com/scholar?as_q=3D-Stacked+1Megapixel+Time-Gated+SPAD+Image+Sensor+with+2D+Interactive+Gating+Network+for+Image+Alignment-Free+Sensor+Fusion&as_occt=title&hl=en&as_sdt=0%2C31)

[^8]: I. Gyongy, A. T. Erdogan, N. A. Dutton, G. M. Martín, A. Gorman, H. Mai, F. M. D. Rocca, and R. K. Henderson, “A Direct Time-of-Flight Image Sensor With In-Pixel Surface Detection and Dynamic Vision,” IEEE Journal of Selected Topics in Quantum Electronics, vol. 30, no. 1 : Single-Photon Technologies and Applications, pp. 1–11, Jan. 2024. [IEEE](https://ieeexplore.ieee.org/document/10023481) [Google Scholar](https://scholar.google.com/scholar?as_q=A+Direct+Time-of-Flight+Image+Sensor+With+In-Pixel+Surface+Detection+and+Dynamic+Vision&as_occt=title&hl=en&as_sdt=0%2C31)

[^9]: R. Gomez-Merchan, J. A. Leñero-Bardallo, R. de la Rosa-Vidal, and Á. Rodríguez-Vázquez, “Dynamic Vision With Single Photon Detectors: A Discrete DVS Architecture Using Asynchronous Sensor Front-Ends,” IEEE Transactions on Circuits and Systems I: Regular Papers, pp. 1–0, 2024. [IEEE](https://ieeexplore.ieee.org/document/10558273) [Google Scholar](https://scholar.google.com/scholar?as_q=Dynamic+Vision+With+Single+Photon+Detectors%3A+A+Discrete+DVS+Architecture+Using+Asynchronous+Sensor+Front-Ends&as_occt=title&hl=en&as_sdt=0%2C31)

[^10]: S. M. Patanwala, I. Gyongy, H. Mai, A. Aßmann, N. A. W. Dutton, B. R. Rae, and R. K. Henderson, “A High-Throughput Photon Processing Technique for Range Extension of SPAD-Based LiDAR Receivers,” IEEE Open Journal of the Solid-State Circuits Society, vol. 2, pp. 12–25, 2022. [IEEE](https://ieeexplore.ieee.org/document/9566373) [Google Scholar](https://scholar.google.com/scholar?as_q=A+High-Throughput+Photon+Processing+Technique+for+Range+Extension+of+SPAD-Based+LiDAR+Receivers&as_occt=title&hl=en&as_sdt=0%2C31)

### Additional References