Search

CN-121995335-A - Millimeter wave radar sparse point cloud diffusion model optimization method

CN121995335ACN 121995335 ACN121995335 ACN 121995335ACN-121995335-A

Abstract

The invention belongs to the technical field of radar signal processing and relates to an optimization method of a millimeter wave radar sparse point cloud diffusion model, which comprises the steps of 1, taking a weighted sum of mean square error and perception loss as a loss function in a baseline model on the basis of a diffusion model, 2, adopting a two-stage strategy, releasing the capacity of generating a structure of the model by adjusting balance points of the loss function on the premise of not sacrificing strong denoising capacity learned by a pre-training model, and 3, constructing an uncertainty model dependent on direction by a radar perception frame through an adaptive error elliptic filter based on radar measurement physical characteristics, and modeling each radar reflection point as an elliptic region with directional error distribution. In particular on the CD and F-score scale, orders of magnitude optimization is achieved.

Inventors

  • LIU JINGTIAN
  • ZHANG SHUZHE
  • ZHANG ZHUXIN
  • XING XIN
  • DING WEN
  • YAO QI
  • YAN ZHIQIANG

Assignees

  • 北方自动控制技术研究所

Dates

Publication Date
20260508
Application Date
20260119

Claims (6)

  1. 1. The millimeter wave radar sparse point cloud diffusion model optimization method is characterized by comprising the following steps of: Step 1, taking a weighted sum of a mean square error and a perceived loss as a loss function in a baseline model on the basis of a Diffusion model Diffusion Models; Step 2, adopting a two-stage Finetuning strategy, and releasing the capability of the model to generate a structure by adjusting the balance point of the loss function on the premise of not sacrificing the learned strong denoising capability of the pre-training model; And 3, constructing a direction-dependent uncertainty model based on radar measurement physical characteristics by the radar perception framework through an adaptive error elliptic filter, and modeling each radar reflection point as an elliptic region with directional error distribution.
  2. 2. The method for optimizing the sparse point cloud diffusion model of the millimeter wave radar according to claim 1, wherein in the step 1, the loss function is: (1) (2) In the formulas (1) and (2), the weights are configured as , In order to calculate the predicted value of the current value, Is true.
  3. 3. The method for optimizing the millimeter wave radar sparse point cloud diffusion model according to claim 2, wherein in the step 2, the two-stage Finetuning strategy specifically comprises: firstly, pre-training, namely pre-training a model by using a baseline model configuration and storing check points; and a second stage of balancing Finetuning, namely loading the check point of the first stage and carrying out cooperative adjustment, namely firstly resetting the optimizer without loading the opt.
  4. 4. The method for optimizing a sparse point cloud model of millimeter wave radar according to claim 3, wherein in the second stage, the weight is adjusted to be And 。
  5. 5. The method for optimizing the sparse point cloud diffusion model of the millimeter wave radar according to claim 1, wherein in the step 3, the final decision criterion of the filter is: For each point in the point cloud First of all use its distance Computing adaptive thresholds And minimum support points Then, the neighborhood points are efficiently searched by constructing KD tree and the support set is calculated The size of (a), i.e ; If and only if the support set size of a point satisfies equation (8), the point is determined to be a structural point and is retained: (8) otherwise, the point is considered an outlier and is culled.
  6. 6. The method for optimizing millimeter wave radar sparse point cloud diffusion model according to claim 5, wherein the threshold value is The method comprises the following steps: (6) Wherein T 0 is the base threshold value, A distance dependent relaxation factor; the minimum support point number The method comprises the following steps: (7) Wherein, the Is the minimum number of support points required near the origin, Is the rate of decay coefficient of the sample, Is a minimum support number lower limit set to prevent the decay from being too fast.

Description

Millimeter wave radar sparse point cloud diffusion model optimization method Technical Field The invention belongs to the technical field of radar signal processing, and particularly relates to a millimeter wave radar sparse point cloud diffusion model optimization method. Background In the complex real world, autonomous navigational capabilities of intelligent systems such as micro-Miniature Aircraft (MAV) and autonomous vehicles are highly dependent on accurate, robust awareness of the surrounding environment. Optical sensors, represented by LiDAR and vision cameras, have become the currently mainstream sensing scheme by virtue of high resolution and high accuracy. However, these optical sensors can drastically reduce or even completely fail in the face of visually degraded environments (Visually Degraded Environments, VDE) such as rain, snow, fog, dust, etc. Scattering and absorption of the light beam lead to sparse point cloud and blurred image, and bring great challenges to safe operation of the system. To overcome this limitation, millimeter wave radar (MMWAVE RADAR) technology is receiving widespread attention in academia and industry due to its unique physical characteristics. Millimeter waves (wavelength is 1-10 mm) of the millimeter wave radar can effectively penetrate through micro particles such as rain, snow and fog, and stable detection in all weather and all days is realized. In addition, millimeter wave radar can also directly measure the speed information (Doppler effect) of a target, making it an ideal supplement or alternative to optical sensors in severe weather. However, despite the all-weather advantages of millimeter wave radar, its application still faces two major core bottlenecks. First, the sparsity of the point cloud is limited by the antenna aperture and bandwidth, and the angular resolution of commercial single chip radars is much lower than that of lidars. This results in an extremely sparse point cloud that it generates, making it difficult to form a fine environmental profile that can be used for localization and mapping. And secondly high noise and multipath interference. Millimeter waves are highly susceptible to multipath reflections when propagating in environments, particularly indoor or canyon environments. This results in the radar detecting a large number of "ghost" points (False Positives), i.e., outliers, at locations other than the true values. The conventional radar signal processing method, such as a Constant False Alarm Rate (CFAR) detector, can extract a target from an original signal, but the output point cloud is sparse and has serious noise, and can not meet the requirement of high-precision navigation. Therefore, a method capable of integrally improving the performance of the millimeter wave radar is needed to solve the above technical problems. Disclosure of Invention The invention provides a millimeter wave radar sparse point cloud diffusion model optimization method, which comprises the following steps: Step 1, taking a weighted sum of a mean square error and a perceived loss as a loss function in a baseline model on the basis of a Diffusion model Diffusion Models; Step 2, adopting a two-stage Finetuning strategy, and releasing the capability of the model to generate a structure by adjusting the balance point of the loss function on the premise of not sacrificing the learned strong denoising capability of the pre-training model; And 3, constructing a direction-dependent uncertainty model based on radar measurement physical characteristics by the radar perception framework through an adaptive error elliptic filter, and modeling each radar reflection point as an elliptic region with directional error distribution. Preferably, in the step 1, the loss function is: (1) (2) In the formulas (1) and (2), the weights are configured as ,In order to calculate the predicted value of the current value,Is true. More preferably, in the step 2, the two-stage Finetuning policy specifically includes: stage one, pre-training, first pre-training a model using a baseline model configuration and saving checkpoints, the model already being The macrostructure and the basic denoising capability of the data are fully learned under the conservation strategy of (1). Stage two, balancing Finetuning, namely loading the check point of stage one and performing cooperative adjustment, namely firstly resetting the optimizer, and not loading the opt_pt state of the pre-training; this allows the model to "tap up" under the gradient of the new loss function, avoiding pulling back the local minimum of the "conservative strategy" by the old optimizer Momentum (Momentum). Second, the learning rate is greatly reduced and fixed at 1e-5, which is a key "stabilizer" to ensure that the model transitions smoothly in the face of new loss function gradients, rather than collapsing due to gradient shocks. More preferably, in the second stage, the weight is adjusted to beAnd. By increasing LPIPS weight