Tracer particle adaptive clustering algorithm based on event camera and speed measurement system

Abstract

The invention discloses a tracer particle adaptive clustering algorithm based on an event camera and a velocity measurement system, and belongs to the field of experimental fluid mechanics flow field velocity measurement. According to the algorithm, continuous laser is used for irradiating flow field tracer particles, and an event camera is used for collecting asynchronous event flows generated by particle movement to obtain event point clouds; particle trajectory reconstruction and velocity field estimation are realized through the steps of statistical filtering, preliminary clustering and velocity fitting, displacement correction, secondary clustering, iterative convergence, velocity field fitting and the like. Microsecond-level time resolution of the event camera is fully utilized, high-frequency laser is replaced by continuous laser, cost is reduced, boundary layer measurement interference is avoided, the problems of time resolution waste and high cost in the prior art are solved, and the method is suitable for experimental fluid mechanics flow field particle tracking speed measurement scenes.

Description

An adaptive clustering algorithm and velocity measurement system for tracer particles based on an event camera Technical Field

[0001] This invention relates to the field of velocity field image processing technology, and in particular to a tracer particle adaptive clustering algorithm and velocity measurement system based on an event camera. Background Technology

[0002] Particle tracking velocimetry (PTV) is a widely used non-contact measurement technique in experimental fluid mechanics, which obtains flow field velocity information by recording the trajectory of tracer particles. With the increasing demand for high-speed flow field measurement, traditional frame-camera-based PTV methods face bottlenecks such as limited frame rate, significant motion blur, high data redundancy, and rising hardware costs, making it difficult to meet the stable measurement requirements of high-speed and low-light scenarios.

[0003] Event cameras, as a novel visual sensor, offer advantages such as microsecond-level temporal resolution, high dynamic range, and low data redundancy, making them suitable for capturing high-speed particle motion. They trigger asynchronous events by independently monitoring changes in light intensity at each pixel, outputting an event stream containing pixel coordinates, timestamps, and brightness polarity, rather than a fixed-frame-rate image. However, existing PTV techniques based on event cameras have significant shortcomings: First, relying on fixed windows to reconstruct pseudo-image frames fails to adapt to the timescales of particles with different velocities, easily leading to trajectory breaks or misalignments, and wasting the high temporal resolution of the event camera; second, the use of high-frequency lasers is not only costly (millions of yuan) but also generates additional event points on the surface of the flow field, making it impossible to measure the flow boundary layer; third, some methods combine with traditional frame cameras, increasing equipment costs and operational complexity. Summary of the Invention

[0004] The core technical reasons for the aforementioned shortcomings are as follows: existing algorithms have not developed dedicated clustering methods for the asynchronous data characteristics of event cameras, and still use the frame-based data processing approach; the laser selection does not match the event camera's light intensity change triggering mechanism, leading to boundary interference; and the equipment combination does not balance cost and performance, limiting engineering applications. Therefore, there is an urgent need for a low-cost PTV algorithm that fully utilizes the advantages of event cameras and adapts to flow field scenarios.

[0005] This invention was made in view of the above-mentioned existing conditions, and its purpose is to provide a tracer particle adaptive clustering algorithm and velocity measurement system based on an event camera. It overcomes the shortcomings of existing event camera-based PTV technology, such as wasted time resolution, high measurement cost, interference with flow boundary layer measurement, and complex operation. It provides a processing algorithm specifically for asynchronous data streams from event cameras. By combining continuous laser and adaptive clustering, it can fully utilize the microsecond-level time resolution of the event camera to achieve high-precision flow field velocity measurement without stacking pseudo-frames, while reducing equipment cost and operation difficulty.

[0006] This invention provides an adaptive clustering algorithm for tracer particles based on an event camera, comprising the following steps: illuminating tracer particles in a flow field with a continuous laser, acquiring an asynchronous event stream generated by the movement of the tracer particles using an event camera to obtain an event point cloud; processing the event point cloud using an adaptive iterative clustering algorithm to achieve particle trajectory reconstruction and velocity field estimation. The processing steps of the adaptive iterative clustering algorithm are as follows: S1: Performing statistical filtering on the initial event point cloud; S2: Performing preliminary clustering on the filtered point cloud to obtain multiple trajectories, and performing linear velocity fitting on each trajectory to obtain the velocity corresponding to each trajectory; S3: Based on the velocity obtained in step S2, calculating the velocity according to formula x... moved =x original - (t present -t start ) * βV performs displacement correction on each point of each trajectory, where β∈[0.6,1.4], x moved Let x be the coordinates of the point after displacement. original Let t be the coordinates of the original point. present t is the timestamp of the current point. start S4: Perform secondary clustering on the displacement-corrected point cloud; S5: Restore the original position of the point cloud; S6: Repeat steps S2-S5 until the number of trajectories converges; S7: Fit the velocity field based on the converged trajectory.

[0007] In this context, the solution reduces equipment costs by combining "continuous laser + event camera" and fully utilizes the microsecond-level temporal resolution of the event camera by avoiding the stacking of pseudo-frames. Adaptive iterative clustering solves the problem of distinguishing trajectories of particles with different velocities through displacement correction and iterative convergence, improving trajectory recognition rate and velocity field estimation accuracy, while avoiding boundary layer measurement interference caused by high-frequency lasers.

[0008] The asynchronous event stream output by the event camera includes pixel position coordinates, timestamps of occurrence, and polarity information of brightness changes.

[0009] This clarifies the core data dimensions of the event stream, providing a data foundation for subsequent timestamp-based displacement correction and trajectory time series analysis, and ensuring that the algorithm can accurately extract the spatiotemporal characteristics of particle motion.

[0010] In step S7, when fitting the velocity field, a time window of 10ms in length centered on the target time is taken, and all trajectories within the time window are linearly fitted to solve for the velocity field at the target time.

[0011] Therefore, the 10ms small time window combined with linear fitting not only utilizes the high temporal resolution of the event camera, but also ensures the accuracy of the velocity field calculation at the target moment through local data fitting, thus achieving accurate velocity measurement under high temporal resolution.

[0012] The continuous laser is used to replace the high-frequency laser.

[0013] Therefore, continuous lasers do not cause sudden changes in light intensity due to high-frequency pulses, which can avoid triggering additional event points on the surface of static objects in the flow field, thus solving the problem of interference in boundary layer measurements in existing technologies. At the same time, the cost of continuous lasers is much lower than that of high-frequency lasers, reducing experimental costs.

[0014] The event camera has a time resolution in the microsecond range.

[0015] Therefore, the microsecond-level temporal resolution enables the algorithm to capture the instantaneous motion state of high-speed particles, avoiding motion blur caused by insufficient frame rate of traditional frame cameras, and providing temporal support for high-precision trajectory reconstruction and velocity calculation.

[0016] In step S1, statistical filtering is used to reduce the noise intensity of the initial event point cloud.

[0017] Therefore, statistical filtering can filter out invalid event points caused by ambient light interference, camera noise, etc., purify the input data, reduce the impact of noise on subsequent clustering accuracy, and improve the stability of trajectory recognition.

[0018] The event point cloud is generated by the light intensity change caused by the movement of tracer particles in the flow field, which triggers the generation of the event camera.

[0019] Therefore, the correlation mechanism between event point clouds and particle motion is clarified, ensuring that the event data processed by the algorithm all originate from the particle motion of the target flow field, thus providing effective data input for the reconstruction of the flow field velocity field.

[0020] In step S2, the preliminary clustering result is at least one trajectory, and each trajectory obtains a unique corresponding velocity through linear fitting.

[0021] This ensures that the velocity characteristics of each trajectory are unique after the initial clustering, providing accurate individual motion parameters for subsequent velocity-based displacement correction and avoiding trajectory correction errors caused by velocity ambiguity.

[0022] This invention also provides a velocity measurement system based on an event camera and an adaptive clustering algorithm for tracer particles. Applying the aforementioned clustering algorithm, the system comprises the following components: an event camera for acquiring flow data to obtain the corresponding event point cloud, which is then imported into the clustering algorithm to obtain the trajectory-fitted velocity field; a continuous laser for illumination, under which the movement of tracer particles generates light intensity changes, and the event camera can capture these light intensity changes to generate an asynchronous event stream containing particle coordinate information; tracer particles, using hollow glass beads, which can follow the flow field and reflect the movement of the transparent fluid; a water tank for containing the flow field; and a thermal convection generator: a resistor is arranged at the bottom of the water tank to heat the resistor and generate upward flow, and a water-cooling device is arranged at the top of the other end of the water tank to reduce downward flow caused by the local water temperature.

[0023] The adaptive clustering algorithm and velocity measurement system for tracer particles based on an event camera provided by this invention have the following advantages compared with the prior art: 1. Significantly improved temporal resolution: The design without stacking pseudo-frames fully utilizes the microsecond-level temporal resolution of the event camera, solving the problem of wasted temporal resolution in existing event camera PTV algorithms; 2. Significantly reduced cost: Continuous laser is used instead of high-frequency laser, eliminating the need for a traditional frame camera, thus significantly reducing equipment cost; 3. Expanded measurement range: The interference from additional event points on the object surface caused by high-frequency laser is avoided, enabling effective measurement of the flow boundary layer; 4. Balance between efficiency and accuracy: The low data volume advantage of the event camera improves computational efficiency, and adaptive iterative clustering improves the accuracy of trajectory recognition and velocity field estimation, with an average velocity error as low as 0.2684 pixels / s and an average coordinate error of only 0.2033 pixels. Attached Figure Description

[0024] Figure 1 shows a schematic diagram of the particle tracking and velocimetry technology based on an event camera-based adaptive clustering algorithm according to an embodiment of the present invention; Figure 2 shows a schematic diagram of the event camera imaging principle of the event camera-based adaptive clustering algorithm according to an embodiment of the present invention; Figure 3 shows a flowchart of the adaptive iterative clustering algorithm of the event camera-based adaptive clustering algorithm according to an embodiment of the present invention; Figure 4 shows a schematic diagram of the principle of utilizing velocity displacement point cloud of the event camera-based adaptive clustering algorithm according to an embodiment of the present invention; Figure 5 shows a comparison diagram of the theoretical velocity field and the velocity field reconstructed by the adaptive iterative clustering algorithm of the event camera-based adaptive clustering algorithm according to an embodiment of the present invention; Figure 6 shows a probability density distribution diagram of the error between the theoretical velocity field and the velocity and particle center coordinates reconstructed by the adaptive iterative clustering algorithm of the event camera-based adaptive clustering algorithm according to an embodiment of the present invention; Figure 7 shows a relative error diagram of the theoretical angular velocity and the measured angular velocity of the disk in the event camera-based adaptive clustering algorithm according to an embodiment of the present invention. Detailed Implementation

[0025] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same parts, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the parts or the shapes of the parts may differ from the actual figures.

[0026] Figure 1 is a schematic diagram of the particle tracking and velocimetry technology based on an event camera-based adaptive clustering algorithm for tracer particles. The system mainly consists of hardware units such as a continuous laser illumination module, tracer particles, and an event camera. The system uses continuous laser light to uniformly illuminate the flow field within the water tank. Under laser illumination, tracer particles (usually hollow glass beads that can effectively follow the flow) moving with the fluid cause rapid changes in local light intensity. The event camera exhibits high temporal resolution and asynchronous response characteristics to these light intensity changes, enabling it to capture the particle motion process in real time and output event point cloud data containing particle spatial location information. The acquired event point cloud data is further input into an adaptive clustering and trajectory analysis algorithm to accurately identify and reconstruct the motion of the tracer particles, ultimately obtaining high temporal resolution full-field velocity field measurement results.

[0027] Figure 2 is a schematic diagram of the imaging principle of an event camera. An event camera is a biomimetic vision sensor whose working principle differs from traditional frame-based cameras. In an event camera, each pixel works independently, continuously monitoring local light intensity changes. When the logarithmic change in light intensity received by a pixel exceeds a preset threshold, an "event" is immediately triggered, and the pixel's spatial coordinates, timestamp, and polarity information are output asynchronously. Therefore, the event camera only responds to dynamic changes in the scene, possessing significant advantages such as microsecond-level temporal resolution, high dynamic range, and low data redundancy.

[0028] This invention addresses the asynchronous event stream output by an event camera by developing a tracer particle adaptive clustering algorithm and velocity field measurement system based on the event camera to achieve particle tracking and velocity measurement. The core objective is to fully utilize the microsecond-level time resolution of the event camera and replace high-frequency lasers with low-cost continuous lasers, thus overcoming the accuracy and cost bottlenecks of existing technologies. The algorithm employs a closed-loop logic of noise reduction, preliminary clustering, velocity correction, secondary clustering, and iterative convergence to accurately extract particle trajectories and reconstruct the velocity field in complex flow fields. The specific process is shown in Figure 3.

[0029] The adaptive clustering algorithm for tracer particles based on an event camera provided in this invention includes: a data acquisition stage: continuous laser irradiation of tracer particles in a flow field, the movement of tracer particles with the flow field causing changes in the intensity of laser reflection, triggering the event camera to output an asynchronous event stream, the event stream containing pixel position coordinates (x, y), occurrence timestamp t, and brightness change polarity, and integrating these events into an initial event point cloud C0 according to the time sequence.

[0030] Adaptive Iterative Clustering Processing: Step 1 Statistical Filtering: Statistical filtering is performed using outlier filtering to reduce noise in the event point cloud, with the number of neighboring points set to 4. This parameter indicates that during noise discrimination, the average distance of each event point is calculated only with its 4 nearest neighbors, and outlier noise points are identified based on the statistical distribution of the average neighborhood distance of the entire point cloud.

[0031] In this implementation, the initial event point cloud (C0) serves as the algorithm's raw input: it is the unprocessed set of events output by the event camera after directly capturing the motion of tracer particles in the flow field and superimposing environmental disturbances. By performing statistical filtering on the initial event point cloud, random noise points are eliminated through neighborhood statistics, taking into account the sparse and noisy event flow characteristics of the event camera output. The core objective is to purify the effective particle event points, providing a reliable data foundation for subsequent iterative clustering, velocity fitting, and trajectory reconstruction; ultimately serving the goal of "high temporal resolution, high precision, and low cost" flow field velocity measurement.

[0032] Step 2: Preliminary Clustering: C1 is segmented using the Euclidean clustering algorithm, with a clustering distance threshold of 2 pixels, resulting in n initial trajectories. Linear fitting is then performed on the time series of each trajectory.

[0033] In some examples, the fitting formulas are x = Vx·t + bx and y = Vy·t + by, calculating the velocity vector V = (Vx, Vy) for each trajectory. Specifically, for the effective event points after noise reduction, Euclidean clustering is used to group them according to spatial-temporal correlation, obtaining several preliminary motion trajectories of particles; then, utilizing the characteristic that "particles are approximately uniform in velocity for a short time," the velocity vector of each trajectory is calculated through linear fitting. This ultimately provides a velocity basis for subsequent splitting of overlapping trajectories and iterative optimization of clustering results.

[0034] Step 3 Displacement Correction: Based on the velocity V of each trajectory, adjust the displacement according to formula x. moved =x original - (t present -t start ) * βVy moved =y original - (t present -t start Coordinate correction is performed using βV, where β can be automatically adjusted within the range [0.6, 1.4] based on the flow field complexity, and t... present t is the timestamp of the current event. start The point cloud C2 is obtained after displacement, using the timestamp of the trajectory start event. This step separates overlapping trajectories by simulating motion compensation viewpoints, as shown in Figure 4.

[0035] Step 4: Secondary Clustering: Using the same Euclidean clustering parameters as in Step 2, C2 is further segmented to obtain a more detailed set of trajectories, resolving the erroneous merging problem caused by trajectory overlap in the initial clustering. In this case, the initial clustering solves the problem of extracting the trajectory outline from the noise; displacement correction solves the problem of indistinguishable overlapping trajectories; and secondary clustering solves the problem of splitting overlapping trajectories. These three steps form a closed loop, ultimately achieving the goal of accurately identifying all particle trajectories and calculating the true velocity. This is a key step supporting the core advantages of this invention: "high recognition rate, high precision, and adaptive flow field."

[0036] Step 5: Coordinate Restoration: The coordinates of the trajectory points after the second clustering are calculated in reverse according to the correction formula to restore them to the original physical space coordinates.

[0037] Step 6 Iterative convergence judgment: Count the number of current trajectories. If the difference between the number of current trajectories and the number of trajectories in the previous iteration is less than 5%, then convergence is determined; otherwise, return to Step 2, refit the velocity based on the new trajectory and repeat the process until convergence.

[0038] Step 7 Velocity Field Fitting: If the velocity field at the target time t0 needs to be solved, take a 10ms time window of [t0-5ms, t0+5ms], perform linear fitting on all converged trajectories within the window to obtain the velocity of each trajectory at time t0, and then generate the full field velocity field distribution through interpolation algorithm.

[0039] In this embodiment, the event point cloud is triggered by changes in light intensity caused by the movement of tracer particles in the flow field, which in turn triggers the generation of an event camera. Event camera generation is a process in which the event camera independently monitors changes in light intensity through pixels, and when the magnitude of the change reaches a certain threshold, outputs event data containing location, time, and polarity in real time.

[0040] In this embodiment, the preliminary clustering result in Step 2 is at least one trajectory, and each trajectory obtains a unique corresponding velocity through linear fitting. This provides the algorithm with a one-to-one quantitative correlation between trajectory and velocity, solving existing technical shortcomings such as "lack of parameter basis for displacement correction, wasted time resolution, and dependence on high-frequency lasers." It also supports the convergence logic of subsequent iterative clustering, ultimately achieving the goal of high time resolution, high precision, low cost, and high efficiency in flow field velocity measurement. This serves as a crucial bridge connecting point cloud preprocessing and accurate trajectory reconstruction, and is the core foundational element enabling the algorithm of this invention to overcome existing technical bottlenecks.

[0041] Experimental verification and results: To verify the effectiveness of the algorithm, two simulation and actual experiments were conducted.

[0042] An open-source flow field simulation software was used to generate a flow field around a cylinder (cylinder diameter 0.02m, incoming flow velocity 20m / s, water medium) to obtain the theoretical velocity field. Following the above implementation steps, an event point cloud was generated and input into the algorithm. The reconstructed velocity field was compared with the theoretical value: 1. Velocity field comparison: As shown in Figure 5, the vector distribution of the reconstructed velocity field is highly consistent with the theoretical velocity field, accurately capturing changes in velocity direction and magnitude even in the vortex shedding region behind the cylinder; 2. Accuracy quantification: As shown in Figure 6, the average velocity error is 0.2684 pixels / s, and the average error of the particle center coordinates is 0.2033 pixels, proving that the algorithm has high measurement accuracy; 3. Performance indicators: The amount of data per second from the event camera depends on the particle concentration and flow field velocity, which is between 60mb / s and 150mb / s. The algorithm takes approximately 1-2 minutes to process 10ms of data on a regular PC, achieving efficient real-time processing. Furthermore, the equipment cost (event camera 45,000 RMB + continuous laser tens of thousands of RMB) is far lower than that of high-frequency laser solutions (hundreds of thousands to millions of RMB).

[0043] To verify the feasibility of the event camera-based tracer particle adaptive clustering algorithm, this experiment designed a spatiotemporal control experiment based on periodic rotational motion. We attached black-based, white-dot targets to a motor-driven turntable, controlling the turntable's rotation speed by adjusting the voltage.

[0044] Under different voltages, an event camera was used to track the trajectory of a white point within 25ms, and the angular velocity was calculated based on the event timestamps and spatial distribution. The calculated angular velocities were then compared with pre-calibrated voltage-speed curves. If the event angular velocities closely matched the calibration results, the feasibility of the algorithm could be verified within the experimental timescale.

[0045] The experimental results are as follows: The relative error of the rotational speed measurement of a tracer particle adaptive clustering algorithm and velocity field measurement system based on an event camera is less than 1.2% under various working conditions, proving that the system has high-precision and high-time-resolution velocity field measurement capabilities.

[0046] As shown in Figure 7, in this embodiment, the relative error between the rotational speed and the theoretical rotational speed was measured at four points with different rotational speeds using a disk experiment. This demonstrates that the event camera-based tracer particle adaptive clustering algorithm and velocity measurement system provided in this embodiment are also accurate and feasible for real experiments.

[0047] The embodiments described above do not constitute a limitation on the scope of protection of this technical solution. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the above embodiments should be included within the scope of protection of this technical solution.

Claims

1. An adaptive clustering algorithm for tracer particles based on an event camera, characterized in that, Includes the following steps: Tracer particles in a flow field are continuously irradiated with laser light, and asynchronous event streams generated by the motion of the tracer particles are collected by an event camera to obtain an event point cloud. An adaptive iterative clustering algorithm is used to process the event point cloud to reconstruct particle trajectories and estimate the velocity field. The processing steps of the adaptive iterative clustering algorithm are as follows: S1: Perform statistical filtering on the initial event point cloud; S2: Perform preliminary clustering on the filtered point cloud to obtain several trajectories, and perform linear velocity fitting on each trajectory to obtain the velocity corresponding to each trajectory; S3: Based on the velocity obtained in step S2, calculate the velocity according to the formula: x moved =x original - (t present -t start ) * βV performs displacement correction on each point of each trajectory, where β∈[0.6,1.4], x moved Let x be the coordinates of the point after displacement. original Let t be the coordinates of the original point. present t is the timestamp of the current point. start S4: Perform secondary clustering on the displacement-corrected point cloud; S5: Restore the original position of the point cloud. S6: Repeat steps S2-S5 until the number of trajectories converges; S7: Fit a velocity field based on the converged trajectory.

2. The event-camera-based adaptive particle clustering algorithm according to claim 1, characterized in that, The asynchronous event stream output by the event camera includes pixel position coordinates, timestamps of occurrence, and polarity information of brightness changes.

3. The event-camera-based adaptive particle clustering algorithm according to claim 1, characterized in that, In step S7, when fitting the velocity field, a time window of 10ms in length is taken with the target time as the center. All trajectories within this time window are linearly fitted to solve for the velocity field at the target time.

4. The event-camera-based adaptive particle clustering algorithm according to claim 1, characterized in that, The continuous laser is used to replace the high-frequency laser.

5. The event-camera-based adaptive particle clustering algorithm according to claim 1, characterized in that, The event camera has a time resolution in the microsecond range.

6. The event-camera-based adaptive particle clustering algorithm according to claim 1, characterized in that, The statistical filtering in step S1 is used to reduce the noise intensity of the initial event point cloud.

7. The event-camera-based adaptive clustering algorithm for tracer particles according to claim 1, characterized in that, The event point cloud is generated by the light intensity change caused by the movement of tracer particles in the flow field, which triggers the generation of the event camera.

8. The event camera-based adaptive clustering algorithm for tracer particles according to claim 1, characterized in that, The result of the preliminary clustering in step S2 is at least one trajectory, and each trajectory obtains a unique corresponding velocity through linear fitting.

9. A velocity measurement system based on an event camera-driven adaptive clustering algorithm for tracer particles, employing the clustering algorithm as described in any one of claims 1-8, characterized in that, The system comprises the following components: an event camera for acquiring flow data, obtaining the corresponding event point cloud, and importing it into the clustering algorithm to obtain the trajectory fitting velocity field; a continuous laser for illumination, under which the movement of tracer particles generates light intensity changes, and the event camera can capture these light intensity changes to generate an asynchronous event stream containing particle coordinate information; tracer particles, which are hollow glass beads, capable of following the flow field and reflecting the movement of the transparent fluid; a water tank for forming and observing the flow field; and a thermal convection generator: a resistor sheet is arranged at the bottom of the water tank, heating the resistor sheet to generate upward flow, and a water cooling device is arranged at the top of the other end of the water tank to reduce the downward flow caused by the local water temperature.

Images