A millimeter wave radar point cloud pedestrian recognition method for an engineering vehicle work scene

Abstract

The application discloses a millimeter wave radar point cloud pedestrian recognition method for an engineering vehicle working scene, and relates to the technical field of safety auxiliary driving of industrial vehicles. The method comprises the following steps: collecting original data of an environment through a vehicle-mounted millimeter wave radar in the first step, performing signal processing on the original data and converting the original data into point cloud information in the second step, clustering the point cloud by using a point cloud clustering algorithm to obtain a target cluster in the third step, and comprehensively analyzing the Doppler speed and variance of the clustered target, a radar cross section (RCS) and spatial position and volume geometric information of the target in the fourth step, so that accurate classification and recognition of pedestrians are realized, and interference targets are filtered out. The method can overcome the influence of light conditions and complex obstacles in an industrial scene, and only millimeter wave radar information is used, so that the anti-interference capability of the radar in a complex industrial environment and the perception capability of the radar for pedestrians are effectively improved while the cost is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of safety assistance driving technology for industrial vehicles, specifically to a millimeter-wave radar point cloud pedestrian recognition method for engineering vehicle operation scenarios. Background Technology

[0002] With the rapid expansion of the global logistics system, engineering vehicles, as core power handling tools in warehousing, manufacturing, and logistics transfer, have reached unprecedented levels in terms of both number and frequency of operation. However, the operating environment of engineering vehicles is highly complex: narrow spaces, frequent human-machine interaction, and numerous blind spots caused by stacked goods have led to engineering vehicle collision accidents (especially those resulting in personal injury) becoming a long-standing problem in the field of industrial safety.

[0003] To reduce accident rates, the industry has introduced various perception technologies. Early safety aids mainly relied on reversing radar (ultrasonic) and physical rearview mirrors, but their short detection range and slow response speed made them ill-suited for dealing with fast-moving pedestrians. Subsequently, pedestrian detection systems based on computer vision (CV) and LiDAR (Light Detection and Ranging) collision avoidance systems began to be applied to mid-to-high-end engineering vehicles. Visual solutions use deep learning models to identify pedestrian outlines, while LiDAR uses high-precision scanning to create environmental maps. Although these technologies have contributed to improving safety, their high cost has limited the widespread deployment of pedestrian recognition technology in engineering vehicles. In the harsh environments of actual industrial applications, the balance between robustness and economy remains elusive.

[0004] Despite the variety of sensing technologies, existing technologies generally suffer from the following deep-seated technical bottlenecks in the specific niche scenario of engineering vehicle operations:

[0005] 1) Environmental Limitations and Computational Redundancy of Visual Perception Solutions: Industrial warehouses and production workshops often experience extreme lighting fluctuations, such as backlit indoor areas, direct sunlight at entrances and exits, and frequent alternations between light and dark. This can lead to severe overexposure or underexposure of camera sensors, resulting in the loss of target features. Furthermore, construction vehicle operations are often accompanied by dust, fumes, and obstructions, significantly reducing the recognition rate of vision systems under these conditions. More importantly, deep learning-based visual recognition places extremely high demands on the computing power of embedded processors in vehicles, and the high hardware cost limits its widespread adoption in ordinary industrial vehicles.

[0006] 2) Reliability and Cost Barriers of LiDAR Solutions: While LiDAR boasts extremely high spatial resolution, its optical components are highly sensitive to dust, requiring frequent maintenance in dusty factory environments. Furthermore, the high-frequency vibrations experienced by engineering vehicles during movement and lifting operations can lead to mechanical fatigue or calibration failure in the precision optical scanning mechanism within the LiDAR under prolonged vibration. More critically, the cost of high-performance LiDAR typically constitutes a significant proportion of the overall cost of the engineering vehicle, lacking widespread market applicability.

[0007] 3) Traditional radar has strong anti-jamming capabilities, but its angle measurement ability is limited, making it difficult to accurately locate targets. Furthermore, factories contain numerous metal shelves, metal equipment, and aluminum-packaged goods. These high RCS (radar cross-section) targets generate extremely complex electromagnetic echoes, even forming "false targets" through multipath reflection, leading to frequent false alarms. Traditional radar typically only provides target distance and speed, lacking in-depth analysis of target shape. In engineering vehicle operations, traditional radar often cannot effectively distinguish between a walking employee and a stationary metal fence or a small metal pillar, causing "alarm fatigue" in drivers due to frequent invalid alarms, ultimately leading them to shut down the safety system and creating potential safety hazards.

[0008] In view of the limitations of the existing technologies in complex industrial scenarios, the development of a pedestrian recognition method that can adapt to strong metal clutter interference, has high reliability in pedestrian classification, and is cost-effective in complex industrial scenarios has become an urgent need in the industry.

[0009] The development of millimeter-wave radar point cloud technology offers a potential solution to this challenge. Unlike traditional radar, millimeter-wave radar (operating in the 30-300 GHz frequency band) can provide richer spatial distribution and micro-Doppler features. With its high-precision ranging, speed sensing, and strong anti-interference capabilities, millimeter-wave radar can adapt well to various scenarios, overcome the limitations of cameras due to external factors, and is also cheaper than lidar. However, how to compensate for motion states in the unique dynamic environment of engineering vehicles, how to extract weak human biometric features from high-intensity metallic noise, and how to achieve a balance between "zero false alarms" and "high detection accuracy" through multi-dimensional logical discrimination remain gaps in the current technological field that urgently need to be addressed. Summary of the Invention

[0010] To address the technical challenges of existing pedestrian detection systems in industrial settings, such as large fluctuations in lighting, severe dust obstruction, strong clutter interference from metal goods and shelves, and target feature shifts caused by the movement of the construction vehicle itself, this invention aims to provide a millimeter-wave radar point cloud pedestrian recognition method for construction vehicle operation scenarios. This method addresses the following specific technical bottlenecks: resolving false alarms in strong metal environments; resolving coordinate consistency issues under dynamic construction vehicle operation; improving the real-time performance of the system on low-computing-power platforms; achieving a lightweight and high-precision point cloud processing workflow; solving the challenge of extracting weak pedestrian features; and accurately locking onto pedestrians through the micro-Doppler effect.

[0011] The specific technical solution for achieving the objective of this invention is as follows:

[0012] A method for pedestrian recognition using millimeter-wave radar point clouds in engineering vehicle operation scenarios includes the following steps:

[0013] Step S1: Radar echo data acquisition and analysis. Pedestrian motion echo signals are acquired using millimeter-wave radar in the engineering vehicle operation scenario, and the initial millimeter-wave radar echo data is obtained through analysis.

[0014] Step S2: Data preprocessing and point cloud dataset generation. The radar echo data collected in step S1 is preprocessed, and motion state compensation and coordinate transformation are performed in combination with the real-time motion state parameters of the engineering vehicle to generate a millimeter-wave radar point cloud dataset.

[0015] Step S3: Clustering of point cloud dataset. The point cloud dataset obtained in step S2 is clustered using a clustering algorithm to identify and extract at least one target cluster.

[0016] Step S4: Multidimensional Feature Extraction and Pedestrian Recognition. Extract the multidimensional feature information of the target clusters obtained in Step S3. This multidimensional feature information includes dynamic speed features, energy distribution features, and geometric morphology features. Based on a multi-constraint discrimination rule, target clusters that meet the preset pedestrian features are identified as pedestrians, and interfering targets in the engineering vehicle operating environment are filtered out. Wherein:

[0017] The motion state compensation in step S2 includes: acquiring the instantaneous linear velocity of the engineering vehicle. and yaw rate The radial velocity of the raw point cloud acquired by radar is corrected in real time, and the true radial compensated velocity of the target relative to the ground is calculated. The target's true radial compensation velocity With radar measurement speed The relationship is:

[0018]

[0019] in, The target azimuth angle, The physical distance from the radar to the rotation center of the engineering vehicle;

[0020] The point cloud dataset is subjected to clustering processing, and the neighborhood radius is used in the clustering process. With detection distance The dynamically compensated adaptive DBSCAN algorithm calculates the neighborhood search radius of distant targets. To compensate for the remote cloud sparseness effect caused by millimeter-wave radar beam divergence, and through The parameters determine whether the point cloud meets the minimum requirements for forming a cluster target. Defined as the minimum point cloud density threshold required to form clusters, ensuring the accuracy of physical boundary segmentation between guardrails, shelves, and pedestrian targets;

[0021]

[0022] in, Based on the basic neighborhood radius, This is the radar resolution compensation coefficient. The radial distance from the point cloud to be clustered to the radar center;

[0023] The multidimensional feature information of the target cluster obtained in extraction step S3 is parsed and fused through the following three parallel methods:

[0024] 1-1. Extract the Doppler velocity variance of each point cloud within the target cluster. By capturing the non-uniform micro-motion characteristics of the limbs relative to the torso during the walking process, the target is identified as a living organism;

[0025] 1-2. Calculate the average radar cross section (RCS) and spatial distribution consistency of the target cluster, and use the physical difference between diffuse reflection from the human body surface and specular reflection from a metal object to filter out high-intensity reflection interference in the engineering vehicle's operating environment.

[0026] 1-3. Extract the 3D bounding box of the target cluster and calculate the height-to-width ratio. The legality of the outline is verified by combining the actual geometric proportions of the pedestrian's body.

[0027] Furthermore, step S1, which involves parsing to obtain the initial millimeter-wave radar echo data, includes: sequentially performing ADC sampling, two-dimensional fast Fourier transform (2D-FFT), and static clutter suppression on the original echo data to obtain initial feature data containing range, velocity, and angle information.

[0028] Furthermore, the radar echo data described in step S2 undergoes data preprocessing, including: performing constant false alarm rate (CFAR) processing and target peak extraction on the original echo signal to extract target points in a complex industrial background.

[0029] Furthermore, the coordinate transformation described in step S2 generates a millimeter-wave radar point cloud dataset. Specific steps include: based on the radar installation height... and radar elevation angle The radar reflection data was transformed from the radar polar coordinate system to the Cartesian coordinate system of the engineering vehicle body using a rotation and translation matrix, and the point cloud coordinates were... Vehicle coordinates :

[0030]

[0031] The generated radar point cloud dataset contains millimeter-wave radar point cloud data with three-dimensional coordinates (x, y, z) as well as Doppler velocity (v) and radar cross-section (RCS).

[0032] Furthermore, the velocity dynamic characteristics mentioned in step S4 include the mean Doppler velocity of each point cloud within the target cluster. and Doppler velocity standard deviation ; Determine whether a target exhibits biological walking characteristics by utilizing fluctuations in Doppler velocity; ... Capture the micro-Doppler effect generated by pedestrian limb movements as a criterion for distinguishing pedestrians from fixed obstacles;

[0033]

[0034] in, The number of points within the target cluster. Let be the Doppler velocity of each point within the target cluster.

[0035] Furthermore, the energy distribution characteristics mentioned in step S4 include the RCS statistics of the target cluster; by combining the RCS statistics with the reflectivity of the target material, and by taking advantage of the difference between the diffuse reflection characteristics of the human body to electromagnetic waves and the specular reflection characteristics of metal shelves and / or vehicles, human targets are distinguished from metal industrial vehicles or shelves; and the RCS distribution characteristics are used to filter out high-intensity metal reflection interference in the working scene of engineering vehicles.

[0036] Furthermore, the geometric features mentioned in step S4 include the bounding box size, volume, and aspect ratio of the target cluster in three-dimensional space; based on the preset pedestrian human body proportion model, the spatial contour of the target cluster is verified, and the projection height of the target cluster is compared with the contour features of common goods and guardrails in the engineering vehicle operation environment as a classification basis.

[0037] Based on the aforementioned speed, energy, and geometric characteristics, a weighted judgment is applied to the target cluster to determine whether the target is a pedestrian. If yes, the pedestrian identification result is output and an alarm signal is triggered; if no, it is determined to be environmental clutter or non-pedestrian interference and filtered out.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] This invention proposes a millimeter-wave radar point cloud-based pedestrian recognition method for engineering vehicle operation scenarios. By combining engineering vehicle motion state compensation, point cloud spatial clustering, and a multi-dimensional feature fusion algorithm incorporating energy distribution and geometric morphology, this invention not only overcomes the influence of ambient lighting and dust but also solves the industry problem of false alarms caused by metallic clutter interference in industrial settings. This has significant practical implications for improving the active safety level of industrial vehicles, reducing enterprise safety management costs, and promoting the intelligent upgrading of logistics equipment.

[0040] This invention, based on the physical characteristics of millimeter-wave radar, effectively overcomes the shortcomings of traditional vision solutions—such as failure under dust and light fluctuations, significant impact from vibration, and interference from industrial metal clutter—by using real-time radial velocity compensation and adaptive clustering of neighborhood radius, combined with micro-Doppler and RCS spatial variance analysis. The system achieves a recognition rate of 95% and a response latency of less than 0.2 seconds, enabling all-weather, highly reliable, and accurate pedestrian perception.

[0041] To address the multipath reflections and clutter interference caused by engineering vehicles operating in confined, high-dynamic environments, this method utilizes motion vectors acquired via the vehicle's CAN bus to perform real-time radial velocity correction on the original point cloud, eliminating observation bias caused by the vehicle's own motion. Simultaneously, by performing multi-dimensional semantic parsing on clustered target clusters, and utilizing the velocity standard deviation... By capturing the unique micro-movement characteristics of pedestrians' limbs and combining them with the spatial distribution patterns of radar cross-section (RCS), the system successfully and accurately locked onto pedestrian targets against the background of highly reflective metal shelves and equipment, solving the technical bottleneck of traditional radar solutions that struggle to distinguish between static metal objects and dynamic living organisms.

[0042] The advantages of this invention lie in its deep integration of "high-precision perception" and "lightweight deployment," enabling it to achieve robustness surpassing traditional vision or LiDAR solutions at a lower hardware cost. Unlike vision solutions' strong dependence on lighting conditions or LiDAR's susceptibility to failure in high-dust, high-vibration environments, the millimeter-wave radar point cloud processing workflow employed in this invention possesses all-weather operation capabilities and utilizes an adaptive neighborhood radius threshold. The consistency of detection at both near and far distances has been optimized, which significantly improves pedestrian recognition accuracy and reduces false alarm rate, while greatly reducing the overall implementation cost of industrial vehicle safety systems. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method of the present invention;

[0044] Figure 2 This is a schematic diagram of the millimeter-wave radar installation of the present invention. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. All other embodiments obtained without other groundbreaking efforts are within the scope of protection of this invention.

[0046] See Figure 1 , Figure 2 This invention relates to a millimeter-wave radar point cloud pedestrian recognition method for engineering vehicle operation scenarios. This method can make full use of the point cloud data detected by millimeter-wave radar to obtain the spatial geometric features, radar reflection features and Doppler velocity features of the target, and determine whether the target belongs to a human based on these features, and output an alarm signal. This greatly controls the cost of pedestrian protection in engineering scenarios and significantly improves the accuracy of pedestrian recognition.

[0047] Example

[0048] The millimeter-wave radar point cloud pedestrian recognition method for engineering vehicle operation scenarios described in this embodiment uses a highly integrated single-chip millimeter-wave radar sensing module as the millimeter-wave radar board. This module integrates a hierarchical phase-locked loop (PLL), a transmitter, a receiver, a baseband processing unit, and an analog-to-digital converter (ADC).

[0049] In this embodiment, forklifts are preferred as a typical application scenario for engineering vehicles. However, in practical applications, forklifts, cranes, and other types of engineering vehicles are not limited to this.

[0050] The preferred operating frequency of the millimeter-wave radar in this embodiment is... In this embodiment, the following is adopted: The radar hardware architecture provides two transmit channels (2TX) and two receive channels (2RX), forming multiple virtual receive channels through multiple transmit-multiple receive (MIMO) technology to support accurate resolution of target azimuth. The preferred output power for a single channel is... This is to ensure that the echo signal has a sufficient signal-to-noise ratio in complex industrial metal reflection environments.

[0051] In this embodiment, the effective identification distance range of the millimeter-wave radar for pedestrians is preferably [missing information]. The preferred horizontal angle measurement range To further illustrate the identification logic and alarm strategy of this invention, this embodiment selects a pedestrian being at a distance from the forklift. At the location, and the radar azimuth angle is at Typical industrial operation scenarios within the scope are described in detail as examples.

[0052] In this embodiment, the millimeter-wave radar is installed above the counterweight at the rear of the forklift, such as... Figure 2 As shown. Preferably, the installation height is... This embodiment uses 1 meter as an example for illustration.

[0053] Figure 2 The diagram illustrates the installation location and detection capabilities of a millimeter-wave radar sensor on an engineering vehicle (using a forklift as an example). Specifically, the millimeter-wave radar sensor 1 is fixedly mounted on the upper rear of the engineering vehicle (e.g., at the rear crossbeam of the overhead guard) via a mounting bracket 2. The mounting bracket 2 is configured with a preset downward tilt angle so that the radar field of view 3 of the millimeter-wave radar sensor 1 can completely cover the ground working area behind the engineering vehicle, thereby achieving effective detection of pedestrians in the blind spot behind the vehicle.

[0054] During operation, the millimeter-wave radar sensor 1 continuously transmits frequency-modulated continuous wave (FMCW) signals into the radar field of view 3 area, and provides raw echo data for subsequent step S1 by receiving echo signals reflected by pedestrians or obstacles.

[0055] It should be understood that the above numerical settings are only to demonstrate the calculation process of this algorithm under specific job constraints, and do not constitute the sole limitation on the scope of protection of this invention.

[0056] The first step involves acquiring data using millimeter-wave radar, with the built-in FMCW Chirp generation engine generating a frequency sweep signal. The radar antenna transmits the signal and receives the echo, which is then sampled at high speed by an internal ADC to convert the radar echo into a digital signal.

[0057] Inside the radar board, the built-in digital signal processing (DSP) unit of the chip directly performs FIR filtering and Fast Fourier Transform (FFT) at the hardware level. This hardware acceleration method greatly reduces the core's computational load, ensuring the system's real-time response when the forklift is traveling at high speed.

[0058] The first dimension FFT is used to extract the target's range information, and the second dimension FFT is used to extract the target's Doppler velocity information, thereby generating a range-Doppler map.

[0059]

[0060] (in The distance from the target to the radar. For Doppler velocity, At the speed of light, The frequency modulation slope reflects the radar's ability to resolve distances. The wavelength of the signal transmitted by the millimeter-wave radar. The frequency difference between the radar's transmitted and received waves at the same moment. (Doppler frequency).

[0061] Spatial filtering algorithms, including but not limited to subtracting two consecutive chirps, are used. The phase of a stationary target remains unchanged, and the subtraction cancels out the chirp; the phase of a moving target changes and is retained, thereby eliminating the signal of stationary fixed obstacles in the background.

[0062] The final result is a raw data cube containing three dimensions: distance, velocity, and angle (azimuth and pitch).

[0063] The second step involves using the CFAR algorithm to dynamically adjust the detection threshold to address the complex electromagnetic noise in forklift operation scenarios, ensuring the stability of target detection under different background noise levels. Furthermore, energy maxima points in the complex scene are extracted for target localization.

[0064]

[0065] (in, The threshold for detection by the Constant False Alarm Rate (CFAR) algorithm. This is the threshold scaling factor. The number of reference units, For the first (Signal power of each reference unit).

[0066] At this point, the forklift motion state compensation calculation is performed in the embedded processor.

[0067] Obtain the instantaneous linear velocity of the forklift itself With angular velocity The azimuth angle of the point cloud target is Calculate the radial compensated velocity of the target relative to the ground. The target's true radial velocity With radar measurement speed The relationship is

[0068]

[0069] (in The distance from the radar's installation location to the center of rotation (the center distance of the radar's installation location) can be used to calculate the object's true velocity. .

[0070] The next step is based on the radar installation height. and radar elevation angle The radar reflection data is transformed from the radar polar coordinate system to the forklift vehicle's Cartesian coordinate system, and the point cloud polar coordinate data measured by the radar is converted. Convert to Cartesian coordinates in the forklift body coordinate system Installation height of millimeter-wave radar Preferred setting is Between, install pitch angle Preferred setting is In this embodiment, the radar installation height is taken as... Pitch angle is (i.e., tilting downwards) (This will be explained.)

[0071]

[0072] The output includes the spatial coordinates of each point. Doppler velocity and radar cross-section The five-dimensional feature points are clustered together.

[0073] The third step is to use the neighborhood radius. With detection distance The dynamically compensated adaptive DBSCAN algorithm aggregates scattered point clouds into independent physical entities and calculates the neighborhood search radius of distant targets. To compensate for the remote cloud sparseness effect caused by millimeter-wave radar beam divergence, and through The parameters determine whether the point cloud meets the minimum requirements for forming a cluster target. Defined as the minimum point cloud density threshold required to form clusters, ensuring the accuracy of physical boundary segmentation between guardrails, shelves, and pedestrian targets;

[0074]

[0075] in, Based on the basic neighborhood radius, This is the radar resolution compensation coefficient. The radial distance from the point cloud to be clustered to the radar center;

[0076] Recursively expand from the core point, merging adjacent points into the same target cluster, while identifying and removing isolated points that cannot form effective clusters as noise.

[0077] The fourth step involves performing in-depth analysis of the target cluster features using a three-way parallel feature extraction module.

[0078] (1) Determine the mean Doppler velocity of each point cloud within the target cluster. and Doppler velocity standard deviation .

[0079] By calculating the mean Doppler velocity of the target cluster To determine its macroscopic movement trend and whether the target possesses the speed and characteristics of a living organism walking.

[0080] Standard deviation of Doppler velocity If the micro-Doppler effect produced by the limb movements of pedestrians is captured, If the value is greater than a preset threshold, the target is determined to have a micro-Doppler effect caused by the swinging of the human limbs, which serves as a criterion for distinguishing pedestrians from fixed obstacles.

[0081]

[0082] in, The number of points within the target cluster. Let be the Doppler velocity of each point within the target cluster.

[0083] (2) The target material is analyzed by calculating the average RCS intensity of the target cluster point cloud and combining it with the spatial distribution characteristics of the RCS. In this embodiment, the system presets the RCS threshold interval for pedestrian recognition as follows: If the average reflection intensity of the target cluster falls within this range, then the spatial characteristics of the target RCS distribution are judged, and only the target cluster point cloud with RCS distribution characteristics that conform to the range of biological radar reflection characteristics is extracted.

[0084] (3) Extract the bounding box size, volume and aspect ratio of the target cluster in three-dimensional space.

[0085] The centroid position of the target cluster is calculated, and the length, width, height, volume, and shape proportion of the target's three-dimensional bounding box are estimated. The spatial contour of the target cluster is verified based on a preset pedestrian human body proportion model. The projection height of the target cluster is compared with the contour features of common goods and guardrails in the engineering vehicle operation environment as a classification basis.

[0086] In this embodiment, the geometric features for determining pedestrians preferably include: a height range of... The width and length range of the bounding box are both The aspect ratio is used as an auxiliary criterion, with the preferred ratio of the target's height to its width being within a certain range. Within the range. In this example, with a height less than Both length and width are less than This serves as the threshold for initial screening of pedestrians. It should be understood that the specific values ​​of the aforementioned geometric features can be flexibly adjusted based on anthropometric data, the wearing of protective gear, and the redundancy of point cloud clustering in the actual work scenario. Furthermore, the geometric features typically need to be combined with dynamic features such as Doppler velocity variance for comprehensive discrimination to filter out non-biological interference targets such as shelves and columns.

[0087] Based on the aforementioned speed, energy, and geometric characteristics, the target cluster is analyzed to determine whether the target is a pedestrian. If yes, the pedestrian identification result is output and an alarm signal is triggered; if no, it is classified as environmental clutter or non-pedestrian interference and filtered out.

[0088] The test consisted of 300 independent randomized trials, covering pedestrians coming from behind the forklift. (Radial) (Oblique) Scenes that enter the detection area from multiple directions, including (crossing) and random directions.

[0089] To simulate a real, complex industrial environment, various typical interference factors were introduced during the test. Goods, office chairs, and other items with strong reflections or multipath reflections were randomly placed in the test area, and tests were conducted in both strong light and low light scenarios.

[0090] The test also examined special scenarios, such as pedestrians pushing metal carts and pedestrians carrying large paper-packaged goods. Under these conditions, the target's RCS energy distribution and geometric profile would change drastically.

[0091] It also added simulation tests for extreme boundary conditions such as pedestrians suddenly appearing from the corner of the shelf and pedestrians standing in front of stationary shelves.

[0092] The final pedestrian recognition rate reached Experiments have shown that, relying on the millimeter-wave radar point cloud pedestrian recognition method for engineering vehicle operation scenarios proposed in this invention, the system can accurately distinguish pedestrians from static objects of similar volume (such as chairs or small goods). Even when pedestrians are holding goods, causing geometric contour distortion, the system can still accurately identify pedestrian characteristics through the micro-Doppler effect.

[0093] In scenarios where only goods and chairs move (simulating goods slipping), the system effectively filters out interference signals through multi-constraint discrimination rules, with only 2 false alarms.

[0094] The average system delay of the forklift alarm signal (from the target entering the detection range to the alarm signal being triggered) stabilizes at Within this range. This specification ensures that the system has sufficient warning time to trigger active braking when the engineering vehicle is traveling at typical operating speeds.

[0095] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for pedestrian recognition using millimeter-wave radar point clouds in engineering vehicle operation scenarios, characterized in that, The method includes the following steps: Step S1: Radar echo data acquisition and analysis. Pedestrian motion echo signals are acquired using millimeter-wave radar in the engineering vehicle operation scenario, and the initial millimeter-wave radar echo data is obtained through analysis. Step S2: Data preprocessing and point cloud dataset generation. The radar echo data collected in step S1 is preprocessed, and motion state compensation and coordinate transformation are performed in combination with the real-time motion state parameters of the engineering vehicle to generate a millimeter-wave radar point cloud dataset. Step S3: Clustering of point cloud dataset. The point cloud dataset obtained in step S2 is clustered using a clustering algorithm to identify and extract at least one target cluster. Step S4: Multidimensional Feature Extraction and Pedestrian Recognition. Extract the multidimensional feature information of the target clusters obtained in Step S3. This multidimensional feature information includes dynamic velocity features, energy distribution features, and geometric morphology features. Based on a multi-constraint discrimination rule, target clusters that meet the preset pedestrian features are identified as pedestrians, and interfering targets in the engineering vehicle operating environment are filtered out. Wherein: The motion state compensation in step S2 includes: acquiring the instantaneous linear velocity of the engineering vehicle. and yaw rate The radial velocity of the raw point cloud acquired by radar is corrected in real time, and the true radial compensated velocity of the target relative to the ground is calculated. The target's true radial compensation velocity With radar measurement speed The relationship is: in, The target azimuth angle, The physical distance from the radar to the rotation center of the engineering vehicle; The multidimensional feature information of the target cluster obtained in extraction step S3 is analyzed and jointly judged through the following three parallel methods: 1-1. Extract the Doppler velocity variance of each point cloud within the target cluster. By capturing the non-uniform micro-motion characteristics of the limbs relative to the torso during the walking process, the target is identified as a living organism; 1-2. Calculate the average radar cross-section and spatial distribution consistency of the target cluster, and use the physical difference between diffuse reflection from the human body surface and specular reflection from a metal object to filter out high-intensity reflection interference in the working environment of the engineering vehicle. 1-3. Extract the 3D bounding box of the target cluster and calculate the height-to-width ratio. The legality of the outline is verified by combining the actual geometric proportions of the pedestrian's body.

2. The method for pedestrian recognition using millimeter-wave radar point cloud in engineering vehicle operation scenarios according to claim 1, characterized in that, Step S1, which involves parsing the initial millimeter-wave radar echo data, includes: sequentially performing ADC sampling, two-dimensional fast Fourier transform, and static clutter suppression on the original echo data to obtain initial feature data containing range, velocity, and angle information.

3. The method for pedestrian recognition using millimeter-wave radar point clouds in engineering vehicle operation scenarios according to claim 1, characterized in that, Step S2 involves preprocessing the radar echo data, including constant false alarm rate (CFAR) processing and target peak extraction of the raw echo signal to extract target points in complex industrial backgrounds.

4. The method for pedestrian recognition using millimeter-wave radar point cloud in engineering vehicle operation scenarios according to claim 1, characterized in that, Step S2 describes coordinate transformation to generate a millimeter-wave radar point cloud dataset. Specific steps include: based on the radar installation height... and radar elevation angle The radar reflection data was transformed from the radar polar coordinate system to the Cartesian coordinate system of the engineering vehicle body using a rotation and translation matrix, and the point cloud coordinates were... Vehicle coordinates : The generated radar point cloud dataset contains millimeter-wave radar point cloud data with three-dimensional coordinates (x, y, z) as well as Doppler velocity (v) and radar cross-section (RCS).

5. The method for pedestrian recognition using millimeter-wave radar point clouds in engineering vehicle operation scenarios according to claim 1, characterized in that, The clustering process described in step S3 uses neighborhood radius during the clustering process. With detection distance The dynamically compensated adaptive DBSCAN algorithm calculates the neighborhood search radius of distant targets. To compensate for the remote cloud sparseness effect caused by millimeter-wave radar beam divergence, and through The parameters determine whether the point cloud meets the minimum requirements for forming a cluster target. Defined as the minimum point cloud density threshold required to form clusters, ensuring the accuracy of physical boundary segmentation between guardrails, shelves, and pedestrian targets; in, Based on the basic neighborhood radius, This is the radar resolution compensation coefficient. The radial distance from the point cloud to be clustered to the radar center is denoted as .

6. The method for pedestrian recognition using millimeter-wave radar point clouds in engineering vehicle operation scenarios according to claim 1, characterized in that, The velocity dynamic characteristics mentioned in step S4 include the mean Doppler velocity of each point cloud within the target cluster. and Doppler velocity standard deviation ; Determine whether a target exhibits biological walking characteristics by utilizing fluctuations in Doppler velocity; ... Capture the micro-Doppler effect generated by pedestrian limb movements as a criterion for distinguishing pedestrians from fixed obstacles; in, The number of points within the target cluster. Let be the Doppler velocity of each point within the target cluster.

7. The method for pedestrian recognition using millimeter-wave radar point clouds in engineering vehicle operation scenarios according to claim 1, characterized in that, The energy distribution characteristics mentioned in step S4 include the RCS statistics of the target cluster; the RCS statistics are combined with the reflectivity of the target material, and the difference between the diffuse reflection characteristics of the human body to electromagnetic waves and the specular reflection characteristics of metal shelves and / or vehicles is used to distinguish human targets from metal industrial vehicles or shelves; the RCS distribution characteristics are used to filter out high-intensity metal reflection interference in the working scene of engineering vehicles.

8. The method for pedestrian recognition using millimeter-wave radar point clouds in engineering vehicle operation scenarios according to claim 1, characterized in that, The geometric features mentioned in step S4 include the bounding box size, volume, and aspect ratio of the target cluster in three-dimensional space; based on the preset pedestrian human body proportion model, the spatial contour of the target cluster is verified, and the projection height of the target cluster is compared with the contour features of common goods and guardrails in the engineering vehicle operation environment as a classification basis.

Images