Method for accurately measuring course angle of unmanned aerial vehicle based on near-infrared laser and reflective pattern

Abstract

The invention discloses an unmanned aerial vehicle course angle accurate measurement method based on near-infrared laser and a reflective pattern, and the method comprises the steps: employing a ground near-infrared imaging system to cooperate with an airborne geometric reflective pattern with the unique directivity, and achieving the measurement through a Vision Transform deep learning model; a large number of label-free images are utilized, random rotation is manually applied, the model predicts the rotation angle, the model obtains the direction characteristics of the reflective pattern under the condition of zero-cost labeling, and then high-precision measurement can be achieved through fine adjustment of a small amount of supervised data. The method solves the problems that the course of the unmanned aerial vehicle is difficult to measure at night and in a weak light environment, and the acquisition cost of labeled data is high, and has the advantages of strong anti-interference capability, simple system deployment and the like.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, specifically relating to a method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns. Background Technology

[0002] Accurate heading angle acquisition for UAVs is crucial for precise navigation, hovering, and complex maneuvering. Common heading angle acquisition methods include: 1) Magnetic compass. This is the cheapest and most common solution, but magnetic compasses are highly susceptible to interference from surrounding ferromagnetic materials and electromagnetic fields, leading to reading drift or even failure. 2) GPS / RTK dual-antenna direction finding. While offering high accuracy, this not only increases the weight and cost of onboard equipment but also fails to function in environments where satellite signals are blocked. 3) Vision-based SLAM or VIO. This method relies on environmental texture features, making it prone to tracking loss at night, in low light, or in scenes with repetitive textures, and it consumes significant onboard computing resources. Currently, there is a lack of a long-distance, high-precision heading measurement method that operates at night or in low light conditions without relying on external satellite signals or consuming excessive onboard computing resources. Summary of the Invention To overcome the shortcomings of existing technologies, this invention provides a method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns. It utilizes a ground-based near-infrared imaging system in conjunction with an airborne geometrically oriented reflective pattern, employing a Vision Transformer deep learning model for measurement. Using a large number of unlabeled images, random rotation is applied manually, and the model predicts the rotation angle, allowing it to learn the directional characteristics of the reflective pattern with zero-cost labeling. High-precision measurement can then be achieved through fine-tuning with a small amount of supervised data. This method solves the problems of difficult UAV heading measurement and high cost of labeled data acquisition in nighttime and low-light environments, and has advantages such as strong anti-interference capability and simple system deployment.

[0003] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Set up a ground-based near-infrared imaging and illumination system; Step 2: Design and arrange the airborne reflective coding pattern; Step 3: Data acquisition and construction of self-supervised pre-training dataset; Step 4: Construction and training of the angle detection model based on Vision Transformer; Step 5: Measure the heading angle accurately online.

[0004] Preferably, step 1 specifically comprises: The ground equipment includes a near-infrared laser, a high-resolution telephoto lens, and an industrial camera. A narrowband filter is installed in front of the telephoto lens. The center wavelength of the narrowband filter is the same as the wavelength of the near-infrared laser to filter out background noise in the natural light and visible light bands, ensuring that only laser reflection signals are received.

[0005] Preferably, step 2 specifically comprises: A ring of omnidirectional reflective strips is attached to the bottom edge of the landing gear or fuselage of the drone. The omnidirectional reflective strips are attached with a unique coded pattern to ensure that the image features presented by the omnidirectional reflective strips at different rotation angles are unique when viewed from the ground, so that the current absolute heading angle of the drone can be deduced from the image.

[0006] Preferably, step 3 specifically comprises: Images of reflective stripes from UAVs were acquired using a ground-based near-infrared imaging and illumination system. A self-supervised learning strategy based on rotation generation was adopted, which uses any image as a reference, applies random rotation and performs zero-filling to generate training samples, and uses the applied rotation angle as a pseudo-label to enable the model to learn and encode the geometric topological features of the pattern as the angle changes.

[0007] Preferably, step 4 specifically comprises: Construct a deep neural network based on Vision Transformer; a) Model architecture: The acquired coded pattern is divided into fixed-size patches, mapped into vector sequences through linear projection, and positional encoding is added; a multi-head self-attention mechanism is used to capture the global context information of the reflective coded strips; b) Training process: divided into two stages; In the first stage, the data collected in step ③ is used for self-supervised pre-training to optimize the encoder parameters; in the second stage, the pre-trained weights are loaded and supervised fine-tuning is performed on the dataset with ground truth angle values. The output layer is changed to a regression head to directly predict the heading angle value θ.

[0008] Preferably, step 5 specifically comprises: The drone is illuminated by a ground-based near-infrared laser, and a telephoto camera captures high-contrast images of reflective stripes through a filter. The images are input into a Vision Transformer-based angle detection model that has been trained, and the model outputs the predicted heading angle in real time. The angle information is then sent to the drone's flight control system via a wireless data transmission link for closed-loop control or to a ground station for display.

[0009] Preferably, the structural features of the coded pattern are as follows: there is no N-order rotational symmetry axis in the two-dimensional plane, where N is greater than or equal to 2, that is, during the process of rotating the pattern around its geometric center from 0 degrees to 360 degrees, it only coincides with the original image at 0 degrees; from a mathematical definition perspective, the geometric centroid of the pattern is forcibly set to not coincide with the center of the smallest circumcircle, and the Euclidean distance D between the two points is greater than a set ratio of the radius of the smallest circumcircle; the vector formed by connecting the center of the circumcircle and the geometric centroid constitutes the unique polar axis direction of the pattern, which is used to physically indicate the absolute heading of the UAV.

[0010] An electronic device includes a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described method for accurately measuring the heading angle of a UAV.

[0011] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for accurately measuring the heading angle of a UAV.

[0012] A chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the aforementioned method for accurately measuring the heading angle of a UAV.

[0013] The beneficial effects of this invention are as follows: This invention utilizes near-infrared narrowband imaging technology, enabling the system to obtain target images with extremely high signal-to-noise ratios even at night and under strong light interference. It employs omnidirectional reflective coding patterns, eliminating the need for onboard active power supply and minimizing the payload of the UAV. By adopting a Vision Transformer model based on self-supervised learning, it effectively overcomes the problem of limited receptive field in traditional convolutional neural networks when processing subtle coding features and solves the problem of scarce labeled data, achieving sub-degree heading measurement accuracy at long distances. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] To address the issues of low accuracy and poor anti-interference capability in heading angle measurement of UAVs under conditions of nighttime, low light, and complex electromagnetic environments, this invention proposes a precise heading angle measurement method for UAVs based on near-infrared laser and reflective patterns. This method combines an active near-infrared light source with passive reflective technology, utilizes optical filtering to eliminate ambient light interference, acquires high signal-to-noise ratio coded images through a telephoto lens, and employs a self-supervised learning-based Vision Transformer model to solve the angle regression problem of small patterns at long distances, achieving all-weather, high-precision external auxiliary measurement.

[0017] The technical solution of the present invention includes the following steps: building a ground near-infrared imaging and illumination system, designing and arranging airborne reflective coding patterns, data acquisition and self-supervised pre-training dataset construction, building and training an angle detection model based on Vision Transformer, and online accurate measurement of heading angle.

[0018] ① Set up a ground-based near-infrared imaging and illumination system: The ground equipment includes a near-infrared laser, a high-resolution telephoto lens, and an industrial camera. A narrow-band filter is installed in front of the telephoto lens. The center wavelength of this filter is the same as the wavelength of the near-infrared laser (within the allowable error range), and the bandwidth is narrow. This is used to filter out background noise in the natural light and visible light bands, ensuring that only the laser reflection signal is received.

[0019] ② Design and layout of airborne reflective coding pattern: A specially designed omnidirectional reflective strip is attached to the bottom edge of the landing gear or fuselage of the UAV. The reflective strip is attached as a unique coded pattern to ensure that the image characteristics presented by the reflective strip at different rotation angles are unique when viewed from the ground, thereby resolving the current absolute heading angle of the UAV from the image.

[0020] ③ Data Acquisition and Self-Supervised Pre-training Dataset Construction: A large number of drone reflective stripe images were collected using the constructed system. To train a robust feature extractor, a rotation-based self-supervised learning strategy was adopted. That is, arbitrary images were used as the reference, random rotations were applied, and zero-padding was performed to generate training samples. The applied rotation angle was used as a pseudo-label, forcing the model to learn the geometric and topological features of the encoded pattern as the angle changes, without relying on manual annotation.

[0021] ④ Construction and training of the angle detection model based on Vision Transformer: Construct a deep neural network based on Vision Transformer.

[0022] a) Model Architecture: The acquired coded pattern image is divided into fixed-size patches, mapped into a vector sequence through linear projection, and positional encoding is added. A multi-head self-attention mechanism is used to capture the global contextual information of the reflective coded strips.

[0023] b) Training process: It is divided into two stages. The first stage uses the data from step ③ for self-supervised pre-training to optimize the encoder parameters; the second stage loads the pre-trained weights and performs supervised fine-tuning on the dataset with ground truth angle values, and changes the output layer to a regression head to directly predict the heading angle value θ.

[0024] ⑤ Online Precise Measurement of Heading Angle: In actual operation, the UAV is illuminated by a ground-based near-infrared laser, and a telephoto camera captures high-contrast reflective stripe images through a filter. The images are input into a trained Vision Transformer angle detection model, which outputs the predicted heading angle in real time. This angle information is then transmitted wirelessly to the UAV flight control system for closed-loop control or to the ground station for display.

[0025] Example: 1. Establish a ground-based near-infrared imaging and illumination system; This invention proposes an optical measurement device deployed at a ground station. A high-power near-infrared laser with a wavelength of 850nm is selected as the active light source; this wavelength is invisible to the human eye and will not cause light pollution. An industrial camera with a telephoto lens of 85mm or longer is used for imaging to ensure clear resolution of the coding details on the bottom of the drone even at a distance. Crucially, a narrow-band interference filter with the same center wavelength as the laser is installed in front of the lens. This filter removes over 90% of ambient light, such as sky background light and streetlights, resulting in an image that retains only bright reflective stripes with an almost completely black background.

[0026] 2. Design and layout of airborne reflective coding patterns; Material selection and arrangement in this invention: In this embodiment, a reflective film material with a high retroreflection coefficient is selected as the marking carrier. Specifically, a high-refractive-index glass microsphere type or microprism type reflective film can be used. To ensure the imaging signal-to-noise ratio at long distances, the retroreflection coefficient of the reflective film should be no less than 300 cd / (lx·m²) under the conditions of an observation angle of 0.2 degrees and an incident angle of -4 degrees. After cutting the reflective material, it is directly pasted onto a flat area on the bottom of the UAV (such as the belly battery compartment cover) or the landing gear bracket to form a reflective area with specific geometric characteristics.

[0027] Geometric Configuration Design Principles: To ensure a unique solution for the heading angle, the reflective coding pattern is designed as a non-centrosymmetric closed geometric figure. The structural characteristic of this pattern is that it lacks an Nth-order rotational symmetry axis (where N is greater than or equal to 2) in the two-dimensional plane; that is, during rotations of the pattern around its geometric center from 0 degrees to 360 degrees, it only coincides with the original image at 0 degrees. From a mathematical definition perspective, the geometric centroid of this pattern is forcibly set to be non-coincident with the center of the smallest circumcircle, and the Euclidean distance D between the two points should be greater than a specific proportion (e.g., 10%) of the radius of the smallest circumcircle. The vector formed by connecting the center of the circumcircle and the geometric centroid constitutes the unique polar axis direction of the pattern, used to physically indicate the absolute heading of the UAV.

[0028] Specific Implementation Configurations: Based on the above principles, this invention provides two preferred pattern configurations. The first is a "concave polygon" design, which is essentially formed by removing a smaller isosceles triangle with the same base from the bottom of an isosceles triangle, creating a concave quadrilateral with a clear directional orientation, its geometric center significantly biased towards the "arrow" head. The second is a "discrete non-uniform dot matrix" design, composed of at least three reflective dots, including a large main positioning point and several auxiliary points. The distribution of the auxiliary points relative to the main positioning point is arranged in a non-isosceles triangular pattern, ensuring that the topological structure formed by the lines connecting the points does not overlap at any rotation angle.

[0029] The coordination mechanism with the measurement model: The projected shape of the pattern designed above on the imaging plane of the ground telephoto camera will undergo a one-to-one perspective geometric transformation as the UAV's heading angle θ changes. During training and inference, the Vision Transformer model extracts the global features of the above asymmetric geometric structure (i.e., the distribution of the shape, the offset of the centroid, and the pointing vector), rather than relying on local texture details, to establish a precise regression mapping relationship between image geometric features and the absolute heading angle θ.

[0030] 3. Data acquisition and self-supervised pre-training based on rotation generation; This invention proposes a self-supervised learning strategy based on a reverse generation mechanism, aiming to address the dependence of deep learning models on large-scale manually labeled data and ensure that the model can accurately capture the topological feature changes of reflective patterns at different angles. The specific implementation process of this method includes the preprocessing of raw data, the construction of synthetic samples, and training logic based on periodic constraints.

[0031] First, the raw data is acquired and standardized preprocessed. Using the aforementioned ground-based near-infrared imaging system, raw image sequences of the UAV in hovering or flight states are acquired. To reduce background noise interference and standardize model input, regions of interest (ROIs) are extracted from the acquired raw images: the images are binarized by setting a brightness threshold to locate the contours of bright reflective areas. The geometric center of this contour is calculated as an anchor point, and the area containing the reflective pattern is cropped into square sub-images of a fixed pixel size (e.g., 224 x 224 pixels). Subsequently, background pixels in the sub-images, except for the main reflective pattern, are uniformly filled with zeros to eliminate ambient stray light interference, forming an unlabeled raw base dataset. At this stage, the system does not need to record the UAV's actual physical heading angle.

[0032] Self-supervised sample construction: For any image in the dataset This is assumed to be a relative "0° reference state". During the training process, for Apply a randomly generated rotation angle The rotated image is generated through affine transformation. .Will As pseudo-labels for the sample, paired training samples are constructed. Pre-training logic: ... Input a ViT model and ask the model to regress and predict the applied rotation angle. By using this "self-testing" approach, the ViT model's attention mechanism is forced to focus on the geometric features in the reflective pattern that change significantly with rotation, rather than the background noise.

[0033] Model pre-training based on periodic loss function: The constructed samples are input into the Vision Transformer model. To address the "numerical abrupt change between zero and 360 degrees" problem in angle regression (i.e., the discontinuity that 359 degrees and 1 degree are physically very close but numerically vastly different), this invention abandons the traditional mean squared error loss and instead uses a periodic loss function for constraint. During training iterations, the model extracts global geometric features from the input image and outputs predicted angle values. The system then calculates the cosine of the difference between the predicted angle value and the true label and subtracts this cosine value from the constant 1 to obtain the final loss value. This trigonometric function-based periodic calculation method ensures that the error calculation remains continuous and small when the predicted value and the true value are located on opposite sides of zero degree, forcing the model to truly understand the rotational topology of the reflective pattern.

[0034] 4. Construction and training of an angle detection model based on Vision Transformer; This invention proposes a deep neural network based on the Vision Transformer, with its output structure specifically modified to adapt to angle regression tasks. Specifically, the classification identifier and its corresponding classification head used in the standard ViT model for classification tasks are removed. Instead, a global average pooling layer is added after the last layer output of the Transformer encoder to compress the feature map into a global feature vector. Subsequently, a multilayer perceptron consisting of two fully connected layers is used as the regression head, with the output dimension set to 1, directly outputting a continuous scalar value representing the predicted heading angle.

[0035] Phase 1: Rotation-Aware Self-Supervised Pre-Training. The model is trained using the synthetic rotation dataset constructed in Step 3. In this phase, some parameters of the model, except for the regression head, are frozen, and the aforementioned "cosine distance periodic loss function" is used as the optimization objective. This phase not only trains the model's sensitivity to geometry but, more importantly, allows the model to learn the topological continuity of the angle space. Phase Two: Supervised Fine-tuning of Real Headings. A small amount of measured data with real heading angle labels is collected. The weights trained in Phase One are loaded, and fine-tuning is performed using the real labels. Since the model has already learned the rules of "relative rotation" during pre-training, this phase requires only a very small number of samples to map relative angles to the absolute geographic coordinate system, achieving high-precision online measurement.

Claims

1. A method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns, characterized in that, Includes the following steps: Step 1: Set up a ground-based near-infrared imaging and illumination system; Step 2: Design and arrange the airborne reflective coding pattern; Step 3: Data acquisition and construction of self-supervised pre-training dataset; Step 4: Construction and training of the angle detection model based on Vision Transformer; Step 5: Measure the heading angle accurately online.

2. The method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns according to claim 1, characterized in that, Step 1 specifically involves: The ground equipment includes a near-infrared laser, a high-resolution telephoto lens, and an industrial camera. A narrowband filter is installed in front of the telephoto lens. The center wavelength of the narrowband filter is the same as the wavelength of the near-infrared laser to filter out background noise in the natural light and visible light bands, ensuring that only laser reflection signals are received.

3. The method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns according to claim 2, characterized in that, Step 2 specifically involves: A ring of omnidirectional reflective strips is attached to the bottom edge of the landing gear or fuselage of the drone. The omnidirectional reflective strips are attached with a unique coded pattern to ensure that the image features presented by the omnidirectional reflective strips at different rotation angles are unique when viewed from the ground, so that the current absolute heading angle of the drone can be deduced from the image.

4. The method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns according to claim 3, characterized in that, Step 3 specifically involves: Images of reflective stripes from UAVs were acquired using a ground-based near-infrared imaging and illumination system. A self-supervised learning strategy based on rotation generation was adopted, which uses any image as a reference, applies random rotation and performs zero-filling to generate training samples, and uses the applied rotation angle as a pseudo-label to enable the model to learn and encode the geometric topological features of the pattern as the angle changes.

5. The method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns according to claim 4, characterized in that, Step 4 specifically involves: Construct a deep neural network based on Vision Transformer; a) Model architecture: The acquired coded pattern is divided into fixed-size patches, mapped into vector sequences through linear projection, and positional encoding is added; a multi-head self-attention mechanism is used to capture the global context information of the reflective coded strips; b) Training process: divided into two stages; In the first stage, the data collected in step ③ is used for self-supervised pre-training to optimize the encoder parameters; in the second stage, the pre-trained weights are loaded and supervised fine-tuning is performed on the dataset with ground truth angle values. The output layer is changed to a regression head to directly predict the heading angle value θ.

6. The method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns according to claim 5, characterized in that, Step 5 specifically involves: The drone is illuminated by a ground-based near-infrared laser, and a telephoto camera captures high-contrast images of reflective stripes through a filter. The images are input into a Vision Transformer-based angle detection model that has been trained, and the model outputs the predicted heading angle in real time. The angle information is then sent to the drone's flight control system via a wireless data transmission link for closed-loop control or to a ground station for display.

7. The method for accurately measuring the heading angle of a UAV based on near-infrared laser and reflective patterns according to claim 6, characterized in that, The structural feature of the coded pattern is that there is no N-order rotational symmetry axis in the two-dimensional plane, where N is greater than or equal to 2. That is, during the process of rotating the pattern around its geometric center from 0 degrees to 360 degrees, it only coincides with the original image at 0 degrees. From a mathematical perspective, the geometric center of the pattern is forced to be non-coincident with the center of the smallest circumcircle, and the Euclidean distance D between the two points is greater than the set ratio of the radius of the smallest circumcircle. The vector formed by connecting the center of the circumcircle and the geometric center constitutes the unique polar axis direction of the pattern, which is used to physically indicate the absolute heading of the UAV.

8. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

10. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 7.

Images