Scale-bias two-parameter self-calibration monocular absolute depth estimation method

Abstract

The application provides a scale-bias double-parameter self-calibration single-frame monocular absolute depth estimation method, mainly improving the precision of single-frame monocular absolute depth estimation and reducing the application cost; the application sets two key parameters of "scale (s)" and "bias (b)" which can be autonomously learned in the network, and uses the geometric relationship between the continuous frames of the video as a "scale" for self-calibration, constantly tries to synthesize the next frame with the depth of the previous frame and the motion of itself, and automatically adjusts the two parameters in the reverse direction by comparing the difference between the synthesized picture and the real picture, so as to form a self-consistent optimization closed loop, and finally drives the depth value output by the network to automatically converge to the real physical scale in "meters", so that the relative depth which is difficult to use directly can be reliably mapped to absolute depth, and the depth estimation precision of the application is significantly improved, and the power consumption cost is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, edge artificial intelligence, unmanned system perception and autonomous navigation, and mainly relates to a single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration. Background Technology

[0002] Monocular depth estimation has long faced the fundamental challenge of scale ambiguity. Existing techniques mainly include traditional geometric methods (such as SFM / MVS), deep learning relative depth estimation methods, multi-sensor fusion methods, and semi-supervised / self-supervised scale recovery methods.

[0003] 1. Traditional geometric methods, such as SFM / MVS, rely on multi-frame feature matching and triangulation, requiring rich textures and a static scene. They are prone to drift in dynamic environments, have a computational latency exceeding 200ms, and cannot output absolute metric depth. Deep learning relative depth estimation methods, such as MiDaS, MonoDepth2, and DPT, output normalized or logarithmic depth, reflecting only relative distance relationships. The scale factor varies with the scene, resulting in actual errors of 1.5–5 meters, making them unsuitable for direct navigation control. Multi-sensor fusion methods, such as VIO and vision-LiDAR, utilize IMU or LiDAR to provide absolute scale, offering high accuracy (<0.3 meters), but increasing hardware costs by 3-10 times, consuming large amounts of power, and being complex to integrate, making them unsuitable for lightweight platforms. Semi-supervised / self-supervised scale recovery methods, such as Depth-from-Video + GPS, rely on continuous video sequences or sparse GPS signals for scale regression, completely failing in hovering, indoor, or signal-obstructed scenes.

[0004] A scale source refers to an external sensor, data, or multi-view geometric constraint that provides absolute physical unit (e.g., meter, centimeter) reference information for vision or other sensor systems. It is crucial for monocular systems with scale uncertainties to recover real-world dimensions. Existing technologies all rely on external scale sources, such as multi-frame, GPS, and LiDAR. Furthermore, current technologies cannot achieve single-frame, single-camera, zero-assistance physical unit depth output, and it is difficult to balance progress, real-time performance, and cost. In summary, current technologies cannot achieve closed-loop output of metric depth from a single camera, resulting in low accuracy. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies and improve the accuracy of single-frame monocular absolute depth estimation, this invention provides a scale-bias dual-parameter self-calibration method for single-frame monocular absolute depth estimation. This invention introduces a globally learnable scale factor *s* and a bias term *b* to achieve closed-loop self-calibration from relative depth to metric absolute depth. It does not rely on external sensors such as GPS, IMU, or LiDAR, or post-processing registration, and is suitable for edge scenarios with stringent requirements for accuracy, real-time performance, and cost, such as UAV altitude holding, autonomous driving obstacle ranging, robot SLAM, and industrial 3D measurement.

[0006] This invention proposes the "dual-parameter closed-loop self-calibration" theory, which breaks through the scale ambiguity barrier of monocular depth estimation and aims to achieve the following goals: (1) pioneering single-camera closed-loop metric output, directly outputting metric absolute depth map based on a single image without external sensors or multi-frame information; (2) high-precision estimation: the average absolute error is less than 1 meter, and about 80% of the pixel errors are less than 1 meter; (3) zero-cost expansion: only a regular RGB camera is required, without the need for an external scale source.

[0007] The technical solution adopted by this invention to solve its technical problem includes the following steps: Step S1: Construct the dataset; The system continuously acquires raw RGB images output from the UAV's downward-looking monocular camera and raw absolute altitudes output from the onboard RTK-GPS module; the acquired raw RGB images and raw absolute altitudes have the same timestamp; the raw RGB images are normalized, and the normalization results are then filtered to obtain RGB images; the width of the RGB images is 1920 pixels, the height of the RGB images is 1080 pixels, and the number of channels in the RGB images is 3; the raw absolute altitudes are cleaned to obtain altitude values; the RGB images and altitude values ​​are used as a set of image data, and multiple sets of image data are acquired to construct a dataset; Step S2: Construct a single-frame monocular absolute depth network; the input of the single-frame monocular absolute depth network is image data; the output of the single-frame monocular absolute depth network is an absolute depth map. The single-frame monocular absolute depth network includes a lightweight feature extraction module, a relative depth prediction module, and a two-parameter closed-loop self-calibration module connected in sequence. The single-frame monocular absolute depth network is trained using a dataset, and the loss function of the single-frame monocular absolute depth network is... After convergence, the optimal single-frame monocular absolute depth network is obtained. Step S3: Normalize the image to be processed to obtain the RGB image of the image to be processed; input the RGB image of the image to be processed and the original absolute height of the image to be processed into the optimal single-frame monocular absolute depth network to obtain the absolute depth map of the image to be processed.

[0008] Furthermore, a one-dimensional Kalman filter is used to process the absolute depth map to obtain a smooth depth map.

[0009] The process of processing the absolute depth map using a one-dimensional Kalman filter to obtain a smooth depth map is as follows: Step 4.1: Construct the state vector State transition matrix Process noise covariance matrix Measurement matrix and measurement noise covariance matrix ; State vector for: ; in, for The absolute depth map d_abs at time step 1. The depth change rate is the rate at which depth changes over time; for each pixel in the absolute depth map. Maintain an independent state vector to describe the evolution of its depth; State transition matrix for: ; in, The sampling time interval is determined by the frame rate of the airborne monocular camera; the state transition matrix... Describe how the system infers the predicted state of the current time from the previous time step, assuming a uniform motion model between adjacent frames: Process noise covariance matrix for: ; Process noise covariance matrix ( ) represents the uncertainty of the motion model itself; Measurement Matrix for: ; Measurement Matrix Because deep networks can only directly observe depth, but cannot directly observe the rate of change, therefore Used to extract depth values ​​from the state vector; Measurement noise covariance matrix for: = 0.5; Measurement noise covariance matrix The noise covariance matrix represents the degree of fluctuation in the output value of the depth estimation network. The reliability of the predicted values ​​and the measured values ​​was weighed; the larger the value, the more trust there is in the prediction model, and the smoother the output. Step 4.2: The process of iteratively calculating and obtaining the smooth depth map is as follows: Step P1: Based on the state vector at time k-1 Error covariance at time k-1 To calculate the prior estimate at the current moment; Step P2: Using the value at time k-1 The predicted value obtained at time k for: That is, predicting the current depth based on the depth and velocity of the previous frame; Based on the error covariance at time k-1 The predicted value of the error covariance at time k is calculated. : ; Step P3: Output the absolute depth map from the dual-parameter closed-loop self-calibration module. As measured value Compared with the predicted value To merge; Calculate Kalman gain : Gain The relative reliability between prediction uncertainty and measurement noise was weighed; Status Update: By correcting the predicted values ​​using measurement residuals, the smoothed optimal depth is obtained. With rate of change ; Covariance update: Update the uncertainty of the current state and return to step P1.

[0010] Furthermore, the lightweight feature extraction module includes a first-stage module Stage1, a second-stage module Stage2, a third-stage module Stage3, a fourth-stage module Stage4, a fifth-stage module Stage5, and a first global average pooling layer, which are cascaded in sequence; the input of the lightweight feature extraction module is image data; the output of the first global average pooling layer is a 512-dimensional global feature vector. The first stage module Stage1 includes a stage convolutional module, a stage max pooling layer, and a first Mish activation function, which are cascaded in sequence; the size of the stage convolutional module is 7×7; the size of the stage max pooling layer is 3×3; The second stage module, Stage2, consists of three cascaded Ghost bottleneck blocks and a second Mish activation function; the second stage module, Stage2, has 128 output channels. The third stage module Stage3 includes four cascaded Ghost bottleneck blocks and a third Mish activation function; the third stage module Stage2 has 256 output channels. The fourth stage module Stage4 includes six cascaded Ghost bottleneck blocks and a fourth Mish activation function; the fourth stage module Stage2 has 512 output channels. The fifth stage module Stage5 includes three cascaded Ghost bottleneck blocks and a fifth Mish activation function; the fifth stage module Stage2 has 512 output channels.

[0011] Furthermore, the relative depth prediction module includes a feature decoding and mapping module and an upsampling recovery module connected in sequence; the input of the relative depth prediction module is a global feature vector of 512; the output of the relative depth prediction module is a depth map. .

[0012] Furthermore, the feature decoding and mapping module includes a first depth module, a second depth module, a third depth module, and a contrastive attention mechanism module connected in sequence; The first depth module includes a fractal-gated aggregation module, a sixth Mish activation function, and a channel-level Dropout module cascaded in sequence; the multi-order gated aggregation module is a Multi-Order Gated Aggregation Module. The dropout probability of the channel-level Dropout module is 20%; The input to the first depth module is a global feature vector of 512; the output of the first depth module is the first enhanced feature; the dimension of the first enhanced feature is 256. The second depth module includes a main path compression module, an auxiliary path compression module, a splicing module, and a contrastive attention mechanism module; the input of the second depth module is a first enhancement feature; the first enhancement feature is input to the main path compression module and the auxiliary path compression module respectively; The main path compression module is a 1×1 convolution kernel; the output of the main path compression module is the main path compression feature, and the dimension of the main path compression feature is 160. The auxiliary path compression module is a depthwise separable convolution; the output of the auxiliary path compression module is an auxiliary path compression feature with a dimension of 96. The splicing module splices the main path compression feature and the auxiliary path compression feature to obtain the module splicing feature; the module splicing feature has 256 dimensions. The contrastive attention mechanism module is an attention mechanism module that integrates the idea of ​​contrastive learning. It enhances the model's discriminative and representational ability of key features by constructing positive and negative sample pairs and comparing their attention responses. The input of the contrastive attention mechanism module is the module concatenation feature. The output of the contrastive attention mechanism module is the dimensionality-reduced feature. The dimension of the dimensionality-reduced feature is 128. The third depth module includes a global depth basis branch module, a local geometry modulation branch module, a coordinate alignment branch module, and a depth fusion module; The input to the third depth module is the dimensionality reduction feature; the dimensionality reduction feature is then input to the global depth basis branch module, the local geometry modulation branch module, and the coordinate alignment branch module, respectively; the output of the third depth module is the initial depth map. The global deep basis branching module includes a global average pooling layer and a global fully connected layer connected in sequence; the input of the global deep basis branching module is the dimensionality reduction feature; the output of the global deep basis branching module is the deep basis function coefficients. The local geometry modulation branch module is a depthwise separable convolution module; the local geometry modulation branch module generates spatially varying geometry modulation factors through lightweight convolution; the output of the local geometry modulation branch module is the geometry modulation factor; The coordinate alignment branch module is a depth-separable convolution module; the output of the coordinate alignment branch module is the offset of each pixel in the x-direction and the offset in the y-direction. The inputs to the deep fusion module include deep basis function coefficients, geometric modulation factors, and the offsets of each pixel in the x and y directions. The deep fusion module process is as follows: Initialize the original depth map; the resolution of the original depth map is 480×270; the original depth map is added pixel by pixel with the offset of each pixel in the x-direction and the offset in the y-direction to obtain the translation depth map; then the translation depth map is multiplied with the geometric modulation factor to obtain the modulation depth map; then the modulation depth map is multiplied with the basis function coefficients to obtain the initial depth map.

[0013] Furthermore, the input to the upsampling recovery module is the initial depth map; the upsampling recovery module uses a bilinear interpolation algorithm to upsample the initial depth map by a factor of 4 to obtain the depth map. .

[0014] Furthermore, the input to the dual-parameter closed-loop self-calibration module is a depth map. The output of the dual-parameter closed-loop self-calibration module is an absolute depth map. The dual-parameter closed-loop self-calibration module for depth maps Perform an affine transformation to obtain the absolute depth map. : ; Where s is a learnable scale factor, initialized to 1.0, which adapts to the camera focal length and scene scale distribution during training; b is a learnable bias term, initialized to 0.0, used to compensate for systematic offsets; This is the final output absolute depth map; the unit of the absolute depth map is meters.

[0015] Furthermore, the loss function for: ; in, Loss of photometric uniformity; For first-order smoothing loss, For scale-regularized loss; The loss of photometric uniformity is fully weighted; For first-order smoothing loss weights, The scaling loss weights are λ1=1.0, λ2=0.1, and λ3=0.001. Loss of photometric uniformity for: ; in It is the set of all valid pixels in an RGB image. This represents the total number of valid pixels. Used to normalize the loss value by average; The coordinates of a specific pixel in an RGB image ; The Structural Similarity Index Measure is an index used to measure the similarity of patches between two RGB images. For the current RGB image At every pixel RGB pixel values ​​at that location; Adjacent frames in an RGB image, i.e. The next frame of the RGB image at that moment; The projection coordinates are the most critical variables in the entire closed-loop calibration; they represent the projection coordinates of the current frame. Pixels Using the absolute depth value predicted by the network And the camera's self-motion between two frames, back-projected into 3D space, and then projected onto the adjacent frame. The new coordinates obtained on the RGB image plane; It is the adjacent frames in the projection coordinates Pixel value at; First-order smoothing loss for: ; First-order smoothing loss Force local smoothing of the depth map while preserving edge structure; among which, These are pixel coordinates; It is a spatial gradient operator used to calculate the spatial derivatives of the depth map in the horizontal and vertical directions, i.e., the differences between adjacent pixels. This is the final metric absolute depth output by a single-frame monocular absolute depth network, calibrated with two parameters. This is the gradient magnitude of the depth map, which measures the degree of depth change between adjacent pixels; It is the input RGB image; This represents the gradient magnitude of the RGB image, i.e., the edge intensity. Edge-aware Weighting; Scale regularization loss for: ; Scale regularization loss To prevent the degradation of s and b, and to maintain physical plausibility; The beneficial effects of this invention are as follows: Compared with existing methods, this invention achieves two improvements: In terms of accuracy and output, this invention can output metric absolute depth without external sensors, reducing the average error to 0.68 meters (about 60% lower than MiDaS and MonoDepth2), and increasing the percentage of pixels with an error of less than 1 meter to 81%; In terms of system and efficiency, this invention completely eliminates the dependence of VIO schemes on IMUs, increasing the inference speed to 43 FPS (existing deep learning-based VIO is only 8 FPS), reducing power consumption to below 5W, and significantly improving energy efficiency. Attached Figure Description

[0016] Figure 1 Here is the flowchart for the timing smoothing module; Figure 2 This is a schematic diagram of the convergence curve during the training process; Figure 3 This is a flowchart of the overall system architecture. Figure 4 Detailed flowchart of the lightweight feature extraction module; Figure 5 Here is a flowchart of the relative depth prediction module; Figure 6 Detailed flowchart of the two-parameter closed-loop self-calibration module; Figure 7 Flowchart for calculating the self-supervised loss function; Figure 8 Here is the flowchart for the timing smoothing module; Figure 9 This is a diagram of the system hardware configuration. Figure 10 This is an error analysis graph; in the graph, Figure 10 (a) is the error distribution histogram; Figure 10 (b) is a distribution chart of predicted and actual values; Figure 10 (c) is a comparison chart of height predictions; Figure 11 The figures show an error bar chart and a cumulative error distribution chart. Figure 11 (a) is a bar chart showing the error distribution; Figure 11 (b) is a cumulative error distribution diagram. Detailed Implementation

[0017] This invention proposes a method for achieving true-scale depth estimation using only monocular video without any external references. This method sets two self-learning key parameters, "scale (s)" and "bias (b)," in the network and uses the geometric relationships between consecutive video frames as a "ruler" for self-calibration: the method continuously attempts to synthesize the next frame using the depth of the previous frame and its own motion, and automatically adjusts the two parameters in reverse by comparing the synthesized image with the real image. This process forms a self-consistent optimization loop, ultimately driving the network's output depth value to automatically converge to the true physical scale in meters, thus reliably mapping relative depth, which is difficult to use directly, to absolute depth.

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] A scale-bias dual-parameter self-calibrated single-frame monocular absolute depth estimation method includes the following steps: Step 0: Build the dataset; The system continuously acquires raw RGB images output from the UAV's downward-looking monocular camera and raw absolute altitude output from the onboard RTK-GPS module; the acquired raw RGB images and raw absolute altitudes have the same timestamp. The original RGB image in the original image is normalized, and the result of the normalization process is then filtered to obtain an RGB image; the width of the RGB image is 1920 pixels, the height of the RGB image is 1080 pixels, and the number of channels of the RGB image is 3; the filtering process is to remove images with motion blur. The original absolute altitude is the altitude value calculated by RTK-GPS; The continuous raw height values ​​are cleaned to obtain height values; the cleaning process is to remove values ​​in the continuous raw height values ​​from which the original absolute height changes abruptly. Collect multiple sets of image data by treating RGB images and height values ​​as a set of image data, and construct a dataset. The dataset used in this invention consists of RGB images captured by a UAV's downward-looking monocular camera and absolute altitude information synchronously output by the onboard RTK-GPS module. To adapt to the fixed input dimension of the deep neural network while preserving sufficient spatial details to recover dense depth, the original captured images are uniformly resized to three-channel RGB images with a width of 1920 pixels and a height of 1080 pixels, i.e., the input tensor shape is 1920×1080×3. The altitude label is in meters, obtained by zeroing the ellipsoidal height or altitude calculated by the RTK-GPS receiver after zeroing the takeoff point altitude, directly reflecting the true vertical distance from the camera to the ground.

[0020] When UAVs fly near the ground or operate in complex environments, RTK-GPS signals are susceptible to multipath effects, poor satellite geometry, or electromagnetic interference. Altitude values ​​in some samples may experience abrupt changes, drift, or even temporary loss, resulting in erroneous labels that severely deviate from the true geometric scale. Directly using this abnormal data for supervised training introduces incorrect scale supervision signals, causing the network to learn an incorrect "image-depth" mapping, severely compromising the accuracy and robustness of absolute depth estimation. Therefore, this invention systematically cleans the original dataset before training. Specifically, it automatically detects and removes frames with obvious RTK-GPS signal anomalies based on the temporal continuity and physical plausibility of altitude values ​​(e.g., altitude changes between adjacent frames exceeding the UAV's maximum ascent / descent rate, or non-positive altitude values). Simultaneously, it filters frames whose image quality is severely degraded due to aircraft shake, motion blur, or occlusion. This data cleaning ensures the consistency and reliability of the labeled altitude with the true physical scale, providing a high-quality supervisory benchmark for the model to accurately recover absolute depth from monocular images. Step 2: Construct a single-frame monocular absolute depth network; The single-frame monocular absolute depth network includes a lightweight feature extraction module and a relative depth prediction module; the single-frame monocular absolute depth network is trained using a dataset, and the loss function of the single-frame monocular absolute depth network is... After convergence, the optimal single-frame monocular absolute depth network is obtained; normalization is applied to the image to be processed to obtain the RGB image of the image to be processed; the RGB image of the image to be processed and the original absolute height of the image to be processed are input into the optimal single-frame monocular absolute depth network to obtain the absolute depth map of the image to be processed. Step 2.1: Construct a lightweight feature extraction module; The lightweight feature extraction module is an improved Ghost-ResNet18 module; The Ghost-ResNet18 improvement module is based on ResNet18 and incorporates the standard features of ResNet18. Replace all convolutional modules with Ghost convolutional modules; The lightweight feature extraction module retains the ability to express multi-scale semantics while compressing the computational cost to 1 / 15 of that of the traditional ResNet50; The input to the lightweight feature extraction module is image data; The output of the lightweight feature extraction module is a 512-dimensional global feature vector; The lightweight feature extraction module includes a first-stage module Stage1, a second-stage module Stage2, a third-stage module Stage3, a fourth-stage module Stage4, a fifth-stage module Stage5, and a first global average pooling layer, which are cascaded in sequence. The first stage module Stage1 includes a stage convolutional module, a stage max pooling layer, and a first Mish activation function, which are cascaded in sequence; the size of the stage convolutional module is 7×7; the size of the stage max pooling layer is 3×3; The second stage module, Stage2, consists of three cascaded Ghost bottleneck blocks and a second Mish activation function; the second stage module, Stage2, has 128 output channels. The third stage module Stage3 includes four cascaded Ghost bottleneck blocks and a third Mish activation function; the third stage module Stage2 has 256 output channels. The fourth stage module Stage4 includes six cascaded Ghost bottleneck blocks and a fourth Mish activation function; the fourth stage module Stage2 has 512 output channels. The fifth stage module Stage5 includes three cascaded Ghost bottleneck blocks and a fifth Mish activation function; the fifth stage module Stage2 has 512 output channels. The output of the first global average pooling layer is a 512-dimensional global feature vector. Throughout the forward propagation of the backbone network from Stage 1 to Stage 5, the Mish activation function is used as the unified nonlinear connection unit after all regular convolutional layers. When a 1920×1080×3 RGB image is input, it first passes through a 7×7 convolutional layer in Stage 1, then undergoes nonlinear mapping using the Mish function, followed by initial downsampling via 3×3 max pooling. Subsequently, it enters multiple Ghost bottleneck blocks stacked in Stages 2 to 5. Within each bottleneck block, whether it is an extended or compressed Ghost Module, the 1×1 convolution and 3×3 depthwise separable convolution are all followed by Mish activation instead of the traditional ReLU. When a skip connection requires channel dimension alignment via a 1×1 convolution, this convolution is also followed by the Mish function. By adopting this step-by-step connection method of "each convolutional layer followed by a Mish activation function", the feature maps can enjoy the characteristics of Mish's smoothness, non-monotonicity and negative gradient flow when they are passed through each stage of the network. This effectively enhances the overall nonlinearity and gradient propagation efficiency of the network in a very lightweight architecture. Finally, a 512-dimensional global feature vector is obtained through global average pooling, which improves the nonlinear expression capability. The lightweight feature extraction module uses depthwise separable Ghost convolutions instead of standard convolutions, reducing FLOPs by 62%; it also introduces channel shuffle to enhance cross-group information flow; and all activation functions are Mish to improve non-linear expressive power.

[0021] Step 2.2: Construct the relative depth prediction module; The core task of the relative depth prediction module is to convert the global feature vector output by the feature extraction module into a depth map that matches the size of the input image and reflects the relative distance between points in the scene. Based on the global feature vector, a multilayer perceptron regression head is constructed to output a single-channel relative depth map (initial resolution 480×270, restored to the original image size through bilinear upsampling).

[0022] The input to the relative depth prediction module is a global feature vector of 512. The output of the relative depth prediction module is a depth map. ; The relative depth prediction module includes a feature decoding and mapping module and an upsampling recovery module connected in sequence. The feature decoding and mapping module includes a first depth module, a second depth module, a third depth module, and a contrastive attention mechanism module connected in sequence. The first depth module includes a fractal-gated aggregation module, a sixth Mish activation function, and a channel-level Dropout module cascaded in sequence; the multi-order gated aggregation module is a Multi-Order Gated Aggregation Module. The dropout probability of the channel-level Dropout module is 20%; The input to the first depth module is a global feature vector of 512; the output of the first depth module is the first enhanced feature; the dimension of the first enhanced feature is 256. The first depth module inputs a 512-dimensional feature tensor into a fractal-gated aggregation module. This module divides the features into four 128-dimensional segments along the channel dimension. Each segment undergoes a non-linear transformation using a miniature multilayer perceptron (128→64→128 neurons). Then, a learnable dynamic gating network (consisting of global average pooling and two fully connected layers) generates independent weights for each segment and performs weighted aggregation accordingly, ultimately outputting a 256-dimensional enhanced feature. Subsequently, the Mish activation function is used, along with channel-level dropout technology, which randomly discards entire feature channels with a 20% probability to promote feature independence. The second depth module includes a main path compression module, an auxiliary path compression module, a splicing module, and a contrastive attention mechanism module; the input of the second depth module is a first enhancement feature; the first enhancement feature is input to the main path compression module and the auxiliary path compression module respectively; The main path compression module is a 1×1 convolution kernel; the output of the main path compression module is the main path compression feature, and the dimension of the main path compression feature is 160. The auxiliary path compression module is a depthwise separable convolution; the output of the auxiliary path compression module is an auxiliary path compression feature with a dimension of 96. The splicing module splices the main path compression feature and the auxiliary path compression feature to obtain the module splicing feature; the module splicing feature has 256 dimensions. The input to the contrastive attention mechanism module is the module concatenation feature; the contrastive attention mechanism module is an attention mechanism module that integrates the idea of ​​contrastive learning, which enhances the model's discriminative and representational ability of key features by constructing positive and negative sample pairs and comparing their attention responses. The output of the contrastive attention mechanism module is a dimensionality-reduced feature; the dimension of the dimensionality-reduced feature is 128. The third depth module includes a global depth basis branch module, a local geometry modulation branch module, a coordinate alignment branch module, and a depth fusion module; The input to the third depth module is the dimensionality reduction feature; the dimensionality reduction feature is then input to the global depth basis branch module, the local geometry modulation branch module, and the coordinate alignment branch module, respectively. The output of the third depth module is the initial depth map; The global deep basis branching module includes a global average pooling layer and a global fully connected layer connected in sequence; the input of the global deep basis branching module is the dimensionality reduction feature; the output of the global deep basis branching module is the deep basis function coefficients. The local geometry modulation branch module is a depthwise separable convolution module; the local geometry modulation branch module generates spatially varying geometry modulation factors through lightweight convolution; the output of the local geometry modulation branch module is the geometry modulation factor; The coordinate alignment branch module is a depth-separable convolution module; the output of the coordinate alignment branch module is the offset of each pixel in the x-direction and the offset in the y-direction. The inputs to the deep fusion module include deep basis function coefficients, geometric modulation factors, and the offsets of each pixel in the x and y directions. The deep fusion module process is as follows: Initialize the original depth map; the resolution of the original depth map is 480×270; the original depth map is added pixel by pixel to the offset of each pixel in the x-direction and the offset in the y-direction to obtain the translation depth map; then the translation depth map is multiplied by the geometric modulation factor to obtain the modulation depth map; then the modulation depth map is multiplied by the basis function coefficients to obtain the initial depth map. These three branches share the same input, namely the 128-dimensional feature map output from the previous layer, but they operate independently and are ultimately fused into an initial depth map through a "neural radiation field heuristic" synthesis logic. Their specific connection relationships are as follows: In the "multi-scale depth field generator", the third layer receives the 128-dimensional feature map output from the previous layer as the sole shared input. This feature map is simultaneously fed into three functionally distinct branches in parallel: The global depth basis branch performs global average pooling on this feature and outputs a set of scalar coefficients through a fully connected layer to control the global "base depth shape" of the entire scene. The local geometry modulation branch performs a lightweight convolution (depthwise separable convolution) operation on the feature, outputting a pixel-wise modulation factor with the same spatial size as the input, which is used to characterize local detail undulations. The coordinate alignment branch also convolves this feature to generate a pixel-level two-dimensional coordinate offset field, used to fine-tune the sampling position to eliminate geometric distortion. Essentially, the coordinate alignment branch module is a process of learning spatial fine-tuning offsets. It receives a shared 128-dimensional feature map and outputs a two-channel coordinate offset field (the offset of each pixel in the x and y directions) through lightweight convolution operations. Its function is to make minor corrections to the standard sampling grid, enabling the depth field to adaptively adjust the sampling position according to the image content during generation and correct geometric distortions caused by preceding network operations. Finally, the outputs of the three branches are synthesized in the spatial dimension: first, the offset field from the coordinate alignment branch adjusts the sampling grid; then, the global depth basis branch determines the depth basis values ​​at various points on the grid; and finally, the factor from the local geometry modulation branch is used for pixel-by-pixel scaling correction, thereby directly generating the final initial low-resolution depth map.

[0023] The input to the upsampling recovery module is the initial depth map; The upsampling recovery module uses a bilinear interpolation algorithm to upsample the initial depth map by a factor of 4 to obtain the depth map. ; The upsampling restoration module restores the input image to the same resolution as the original input image, thereby obtaining a depth map where each pixel has a depth value. ; Step 2.3: Construct a two-parameter closed-loop self-calibration module; The input to the dual-parameter closed-loop self-calibration module is a depth map. ; The output of the dual-parameter closed-loop self-calibration module is an absolute depth map. ; The two-parameter closed-loop self-calibration module for depth maps Perform an affine transformation to obtain the absolute depth map. : ; Where s is a learnable scale factor, initialized to 1.0, which adapts to the camera focal length and scene scale distribution during training; b is a learnable bias term, initialized to 0.0, used to compensate for systematic offsets (such as camera mounting height). This is the final output metric absolute depth; Step 3: Construct the loss function : ; in, Loss of photometric uniformity; For first-order smoothing loss, For scale-regularized loss; The loss of photometric uniformity is fully weighted; For first-order smoothing loss weights, The scaling loss weights are λ1=1.0, λ2=0.1, and λ3=0.001. Photometric consistency loss is constructed by utilizing the photometric invariance between consecutive frames. for: ; in It is the set of all valid pixels in an RGB image, usually excluding pixels that are occluded or exceed the boundaries. This represents the total number of valid pixels. Used to normalize the loss value by average; The coordinates of a specific pixel in an RGB image ; The Structural Similarity Index Measure is an index used to measure the similarity of patches between two RGB images. For the current video frame At every pixel RGB pixel values ​​at that location; Adjacent frames in a video sequence, i.e. The next frame of the moment; The projection coordinates are the most critical variables in the entire closed-loop calibration; they represent the projection coordinates of the current frame. Pixels Using the absolute depth value predicted by the network And the camera's self-motion between two frames, back-projected into 3D space, and then projected onto the adjacent frame. The new coordinates obtained on the RGB image plane; It is the adjacent frames in the projection coordinates The pixel value at that location is "based on the estimated depth and motion, the corresponding pixel in the next frame is 'cut out' and pasted onto the current frame's position." If the depth and motion estimation is completely accurate, the synthesized pixel should be consistent with... Highly consistent; First-order smoothing loss for: ; First-order smoothing loss Force local smoothing of the depth map while preserving edge structure; among which, These are pixel coordinates; It is a spatial gradient operator used to calculate the spatial derivatives of the depth map in the horizontal and vertical directions, i.e., the differences between adjacent pixels. This is the final metric absolute depth output by a single-frame monocular absolute depth network, calibrated with two parameters. This is the gradient magnitude of the depth map, which measures the drasticness of depth changes between adjacent pixels. During training of a single-frame monocular absolute depth network, this value is suppressed as much as possible to force the depth map to become smooth. It is the input RGB image; This represents the gradient magnitude of the RGB image, i.e., the edge intensity. This is the edge-aware weighting term; it's a weight that dynamically changes with the image gradient. In flat areas, the image color is uniform, like a flat ground viewed from above. As the exponential term approaches 0, the depth gradient approaches 1. It will be severely penalized, forcing the depth of that area to remain continuously smooth; at the edges of objects where the image color changes abruptly, such as at the boundary between a moving object and the background ground, The value is very large, and the exponent term approaches 0, which is equivalent to "turning off" the penalty for depth changes at this location, allowing the depth to undergo reasonable physical jumps at the edge of the target's outline, thereby accurately segmenting individual targets; Scale regularization loss for: ; Scale regularization loss To prevent the degradation of s and b, and to maintain physical plausibility; Step 4: Process the absolute depth map using a one-dimensional Kalman filter to obtain a smooth depth map; Step 4.1: Construct the state vector State transition matrix Process noise covariance matrix Measurement matrix and measurement noise covariance matrix ; State vector for: ; in, for The absolute depth map d_abs at time step 1. The depth change rate is the rate at which depth changes over time; for each pixel in the absolute depth map. Maintain an independent state vector to describe the evolution of its depth; State transition matrix for: ; in, The sampling time interval is determined by the frame rate of the airborne monocular camera; the state transition matrix... Describe how the system infers the predicted state of the current time from the previous time step, assuming a uniform motion model between adjacent frames: Process noise covariance matrix for: ; Process noise covariance matrix ( ) represents the uncertainty of the motion model itself; Measurement Matrix for: ; Measurement Matrix Because deep networks can only directly observe depth, but cannot directly observe the rate of change, therefore Used to extract depth values ​​from the state vector; Measurement noise covariance matrix for: = 0.5; Measurement noise covariance matrix The noise covariance matrix represents the degree of fluctuation in the output value of the depth estimation network. The reliability of the predicted values ​​and the measured values ​​was weighed; the larger the value, the more trust there is in the prediction model, and the smoother the output. Step 4.2: The process of iteratively calculating and obtaining the smooth depth map is as follows: Step P1: Based on the state vector at time k-1 Error covariance at time k-1 To calculate the prior estimate at the current moment; Step P2: Using the value at time k-1 The predicted value obtained at time k for: That is, predicting the current depth based on the depth and velocity of the previous frame; Based on the error covariance at time k-1 The predicted value of the error covariance at time k is calculated. : ; Step P3: Output the absolute depth map from the dual-parameter closed-loop self-calibration module. As measured value Compared with the predicted value To merge; Calculate Kalman gain : Gain The relative reliability between prediction uncertainty and measurement noise was weighed; Status Update: By correcting the predicted values ​​using measurement residuals, the smoothed optimal depth is obtained. With rate of change ; Covariance update: Update the uncertainty of the current state and return to step P1; After one-dimensional Kalman filtering, a smooth depth map is obtained; each pixel value in the smooth depth map is the actual physical distance after temporal denoising. To suppress dynamic jitter, a one-dimensional Kalman filter is introduced to perform temporal smoothing on each pixel depth. The flowchart of the temporal smoothing and output module is as follows. Figure 1 As shown, the Kalman filter performs prediction and updates through step 4.2: iterative calculation to obtain a smooth depth map. The one-dimensional Kalman filter process is performed independently pixel-by-pixel and frame-by-frame, ultimately outputting a 1920×1080×1 meter depth map, where each pixel value represents the actual physical distance (in meters) from the camera to the scene point.

[0024] Judging the convergence effectiveness of single-frame monocular absolute depth network training process: This invention can drive the spontaneous convergence of global network parameters solely by the geometric consistency constraints of multi-frame sequences, without external absolute scale truth supervision.

[0025] scale factor Macroscopic fast locking: parameters Starting with an initial value of 1.0, it is extremely sensitive to the loss gradient in the early stages of training, exhibiting a rapid convergence trend. The core physical significance of this stage lies in quickly capturing and locking the camera's focal length characteristics and the overall depth distribution of the scene, thereby achieving rapid macroscopic spatial alignment from the normalized relative depth domain to the physical metric depth domain.

[0026] Bias term Microscopic fine compensation: parameters Starting with an initial value of 0.0, its convergence trajectory is relatively smooth. At the scaling factor... During or after the process of stabilization, the parameters Gradually make small, fine-tuning adjustments. The physical purpose is to accurately compensate for systematic static deviations such as the fixed installation height of the UAV's downward-facing camera.

[0027] Without human intervention, it can automatically find the optimal value and converge to the physical true value, verifying the theoretical and engineering feasibility of the "dual-parameter closed-loop self-calibration" mechanism of this invention.

[0028] Overview of the scenario for this invention embodiment: This embodiment addresses the need for precise altitude hold during unmanned aerial vehicle (UAV) operations such as patrol flights, requiring a fixed altitude above the ground. Traditional methods rely on barometers or ultrasonic sensors; the former is susceptible to weather-related drift, while the latter has a limited effective range. This invention utilizes a scale-bias dual-parameter self-calibrated monocular absolute depth estimation method, employing only a downward-viewing monocular camera, to achieve metric-level accuracy in altitude measurement, effectively overcoming the inherent limitations of traditional sensors.

[0029] System Hardware Configuration: The hardware platform used in this invention is extremely streamlined, perfectly meeting the requirements of lightweight and low-cost drones. An NVIDIA Jetson Orin NX embedded computing unit serves as the onboard processing core; A downward-looking global shutter RGB camera (model Sony IMX545) is fixed to the belly of the drone and mounted vertically downwards to capture ground images at a resolution of 1920×1080. A Pixhawk 6C flight control system communicates with Jetson Orin NX via UART serial port to receive altitude commands.

[0030] The system hardware configuration diagram is as follows: Figure 9 As shown.

[0031] System software framework: The system algorithm was implemented in an Ubuntu 20.04 environment, using the Noetic version of the Robot Operating System (ROS). The core model of this invention—a scale-bias dual-parameter self-calibrating depth estimation network—was accelerated using TensorRT and deployed on the Jetson platform, encapsulated as an independent ROS service node that receives images and returns absolute depth values.

[0032] The specific implementation steps are as follows: 1: Sensor data acquisition and dissemination; After the drone takes off, the downward-looking monocular camera continuously captures ground images directly below the drone's belly at a frequency of 30Hz and publishes them in real time via the ROS topic / downward / image_raw.

[0033] 2: Depth estimation service nodes; This step periodically calls the depth estimation service node to perform forward inference on each incoming frame of image. The model maps the relative depth output by the network to the absolute depth using an internally globally learnable scale factor s and a bias term b, and extracts the median depth of the central region of the image (such as a 10×10 pixel block) as the real-time ground clearance d_center.

[0034] 3: Calculate the height deviation; In this step, the altitude control module in the flight control system subscribes to the depth topic and continuously reads d_center. This module compares d_center with the preset target altitude (500.0 meters) and calculates the altitude deviation e = d_center - 500.0.

[0035] 4: Height Adjustment This step uses a PID controller to perform closed-loop calculations on the altitude deviation e, generating corresponding throttle control quantities to counteract altitude fluctuations and stabilize the drone at the target altitude.

[0036] 5: Highly autonomous, stable, and controllable In this step, the flight control system sends the final control command to the ESC to precisely adjust the motor speed, achieving autonomous, continuous, and stable altitude maintenance.

[0037] The test results are as follows Figure 10 and Figure 11 As shown.

[0038] Comparison of the effects of this invention with existing technologies:

[0039] Compared to existing methods, this invention achieves two improvements: In terms of accuracy and output, this invention can output metric absolute depth without external sensors, reducing MAE to 0.68 meters and increasing the percentage of pixels with an error of less than 1 meter to 81%; In terms of system and efficiency, this invention completely eliminates the dependence of VIO solutions on IMUs, increasing inference speed to 43 FPS (compared to only 8 FPS for existing deep learning-based VIO), reducing power consumption to below 5W, and significantly improving energy efficiency; The core technical advantage of this invention lies in achieving absolute scale closed-loop recovery with only two parameters, and its strong engineering feasibility: significantly improved inference speed, suitable for native deployment on edge devices. Based on a single-camera solution, the total cost is lower than depth estimation methods such as multi-sensor fusion and LiDAR, possessing the potential to replace LiDAR.

Claims

1. A single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration, characterized in that, Includes the following steps: Step S1: Construct the dataset; The system continuously acquires raw RGB images output from the UAV's downward-looking monocular camera and raw absolute altitude output from the onboard RTK-GPS module; the acquired raw RGB images and raw absolute altitudes have the same timestamp. The original RGB image in the original image is normalized, and the result of the normalization process is then filtered to obtain the RGB image; the width of the RGB image is 1920 pixels, the height of the RGB image is 1080 pixels, and the number of channels of the RGB image is 3. The original absolute height is cleaned to obtain the height value; the RGB image and the height value are used as a set of image data, and multiple sets of image data are collected to construct a dataset; Step S2: Construct a single-frame monocular absolute depth network; the input of the single-frame monocular absolute depth network is image data; the output of the single-frame monocular absolute depth network is an absolute depth map. The single-frame monocular absolute depth network includes a lightweight feature extraction module, a relative depth prediction module, and a two-parameter closed-loop self-calibration module connected in sequence. The single-frame monocular absolute depth network is trained using a dataset, and the loss function of the single-frame monocular absolute depth network is... After convergence, the optimal single-frame monocular absolute depth network is obtained. Step S3: Normalize the image to be processed to obtain the RGB image of the image to be processed; input the RGB image of the image to be processed and the original absolute height of the image to be processed into the optimal single-frame monocular absolute depth network to obtain the absolute depth map of the image to be processed.

2. The single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration according to claim 1, characterized in that, The process of processing the absolute depth map using a one-dimensional Kalman filter to obtain a smooth depth map is as follows: Step 4.1: Construct the state vector State transition matrix Process noise covariance matrix Measurement matrix and measurement noise covariance matrix ; State vector for: ； in, for The absolute depth map d_abs at time step 1. The depth change rate is the rate at which depth changes over time; for each pixel in the absolute depth map. Maintain an independent state vector to describe the evolution of its depth; State transition matrix for: ； in, The sampling time interval is determined by the frame rate of the airborne monocular camera; the state transition matrix... Describe how the system infers the predicted state of the current time from the previous time step, assuming a uniform motion model between adjacent frames: Process noise covariance matrix for: ； Process noise covariance matrix ( ) represents the uncertainty of the motion model itself; Measurement Matrix for: ； Measurement Matrix Because deep networks can only directly observe depth, but cannot directly observe the rate of change, therefore Used to extract depth values ​​from the state vector; Measurement noise covariance matrix for: = 0.5; Measurement noise covariance matrix The noise covariance matrix represents the degree of fluctuation in the output value of the depth estimation network. The reliability of the predicted values ​​and the measured values ​​was weighed; the larger the value, the more trust there is in the prediction model, and the smoother the output. Step 4.2: The process of iteratively calculating and obtaining the smooth depth map is as follows: Step P1: Based on the state vector at time k-1 Error covariance at time k-1 To calculate the prior estimate at the current moment; Step P2: Using the value at time k-1 The predicted value obtained at time k for: That is, predicting the current depth based on the depth and velocity of the previous frame; Based on the error covariance at time k-1 The predicted value of the error covariance at time k is calculated. : ； Step P3: Output the absolute depth map from the dual-parameter closed-loop self-calibration module. As measured value Compared with the predicted value To merge; Calculate Kalman gain : Gain The relative reliability between prediction uncertainty and measurement noise was weighed; Status Update: By correcting the predicted values ​​using measurement residuals, the smoothed optimal depth is obtained. With rate of change ; Covariance update: Update the uncertainty of the current state and return to step P1.

3. The single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration according to claim 1, characterized in that, The lightweight feature extraction module includes a first-stage module Stage1, a second-stage module Stage2, a third-stage module Stage3, a fourth-stage module Stage4, a fifth-stage module Stage5, and a first global average pooling layer, which are cascaded in sequence. The input of the lightweight feature extraction module is image data. The output of the first global average pooling layer is a 512-dimensional global feature vector. The first stage module Stage1 includes a stage convolutional module, a stage max pooling layer, and a first Mish activation function, which are cascaded in sequence; the size of the stage convolutional module is 7×7; the size of the stage max pooling layer is 3×3; The second stage module, Stage2, consists of three cascaded Ghost bottleneck blocks and a second Mish activation function; the second stage module, Stage2, has 128 output channels. The third stage module Stage3 includes four cascaded Ghost bottleneck blocks and a third Mish activation function; the third stage module Stage2 has 256 output channels. The fourth stage module Stage4 includes six cascaded Ghost bottleneck blocks and a fourth Mish activation function; the fourth stage module Stage2 has 512 output channels. The fifth stage module Stage5 includes three cascaded Ghost bottleneck blocks and a fifth Mish activation function; the fifth stage module Stage2 has 512 output channels.

4. The single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration according to claim 1, characterized in that, The relative depth prediction module includes a feature decoding and mapping module and an upsampling recovery module connected in sequence; the input of the relative depth prediction module is a global feature vector of 512; the output of the relative depth prediction module is a depth map. .

5. The single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration according to claim 4, characterized in that, The feature decoding and mapping module includes a first depth module, a second depth module, a third depth module, and a contrastive attention mechanism module connected in sequence. The first depth module includes a fractal-gated aggregation module, a sixth Mish activation function, and a channel-level Dropout module cascaded in sequence; the multi-order gated aggregation module is a Multi-Order Gated Aggregation Module. The dropout probability of the channel-level Dropout module is 20%; The input to the first depth module is a global feature vector of 512; the output of the first depth module is the first enhanced feature. The dimension of the first enhancement feature is 256; The second deep module includes a main path compression module, an auxiliary path compression module, a concatenation module, and a contrastive attention mechanism module; the input of the second deep module is the first enhanced feature; The first enhanced features are input into the main path compression module and the auxiliary path compression module, respectively; The main path compression module is a 1×1 convolution kernel; the output of the main path compression module is the main path compression feature, and the dimension of the main path compression feature is 160. The auxiliary path compression module is a depthwise separable convolution; the output of the auxiliary path compression module is an auxiliary path compression feature with a dimension of 96. The splicing module splices the main path compression feature and the auxiliary path compression feature to obtain the module splicing feature; the module splicing feature has 256 dimensions. The contrastive attention mechanism module is an attention mechanism module that integrates the idea of ​​contrastive learning. It enhances the model's discriminative and representational ability of key features by constructing positive and negative sample pairs and comparing their attention responses. The input of the contrastive attention mechanism module is the module concatenation feature. The output of the contrastive attention mechanism module is the dimensionality-reduced feature. The dimension of the dimensionality-reduced feature is 128. The third depth module includes a global depth basis branch module, a local geometry modulation branch module, a coordinate alignment branch module, and a depth fusion module; The input to the third depth module is the dimensionality reduction feature; the dimensionality reduction feature is then input to the global depth basis branch module, the local geometry modulation branch module, and the coordinate alignment branch module, respectively; the output of the third depth module is the initial depth map. The global deep basis branching module includes a global average pooling layer and a global fully connected layer connected in sequence; the input of the global deep basis branching module is the dimensionality reduction feature; the output of the global deep basis branching module is the deep basis function coefficients. The local geometry modulation branch module is a depthwise separable convolution module; the local geometry modulation branch module generates spatially varying geometry modulation factors through lightweight convolution; the output of the local geometry modulation branch module is the geometry modulation factor; The coordinate alignment branch module is a depth-separable convolution module; the output of the coordinate alignment branch module is the offset of each pixel in the x-direction and the offset in the y-direction. The inputs to the deep fusion module include deep basis function coefficients, geometric modulation factors, and the offsets of each pixel in the x and y directions. The deep fusion module process is as follows: Initialize the original depth map; the resolution of the original depth map is 480×270; the original depth map is added pixel by pixel with the offset of each pixel in the x-direction and the offset in the y-direction to obtain the translation depth map; then the translation depth map is multiplied with the geometric modulation factor to obtain the modulation depth map; then the modulation depth map is multiplied with the basis function coefficients to obtain the initial depth map.

6. The single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration according to claim 4, characterized in that, The input to the upsampling recovery module is the initial depth map; the upsampling recovery module uses a bilinear interpolation algorithm to upsample the initial depth map by a factor of 4 to obtain the depth map. .

7. The single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration according to claim 1, characterized in that, The input to the dual-parameter closed-loop self-calibration module is a depth map. The output of the dual-parameter closed-loop self-calibration module is an absolute depth map. The dual-parameter closed-loop self-calibration module for depth maps Perform an affine transformation to obtain the absolute depth map. : ； Where s is a learnable scale factor, initialized to 1.0, which adapts to the camera focal length and scene scale distribution during training; b is a learnable bias term, initialized to 0.0, used to compensate for systematic offsets; This is the final output absolute depth map; the unit of the absolute depth map is meters.

8. The single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration according to claim 1, characterized in that, The loss function for: ； in, Loss of photometric uniformity; For first-order smoothing loss, For scale-regularized loss; The loss of photometric uniformity is fully weighted; For first-order smoothing loss weights, The scaling loss weights are λ1=1.0, λ2=0.1, and λ3=0.

001. Loss of photometric uniformity for: ； in It is the set of all valid pixels in an RGB image. This represents the total number of valid pixels. Used to normalize the loss value by average; The coordinates of a specific pixel in an RGB image ; The Structural Similarity Index Measure is an index used to measure the similarity of patches between two RGB images. For the current RGB image At every pixel RGB pixel values ​​at that location; Adjacent frames in an RGB image, i.e. The next frame of the RGB image at that moment; The projection coordinates are the most critical variables in the entire closed-loop calibration; they represent the projection coordinates of the current frame. Pixels Using the absolute depth value predicted by the network And the camera's self-motion between two frames, back-projected into 3D space, and then projected onto the adjacent frame. The new coordinates obtained on the RGB image plane; It is the adjacent frames in the projection coordinates Pixel value at; First-order smoothing loss for: ； First-order smoothing loss Force local smoothing of the depth map while preserving edge structure; among which, These are pixel coordinates; It is a spatial gradient operator used to calculate the spatial derivatives of the depth map in the horizontal and vertical directions, i.e., the differences between adjacent pixels. This is the final metric absolute depth output by a single-frame monocular absolute depth network, calibrated with two parameters. This is the gradient magnitude of the depth map, which measures the degree of depth change between adjacent pixels; It is the input RGB image; This represents the gradient magnitude of the RGB image, i.e., the edge intensity. Edge-aware Weighting; Scale regularization loss for: ； Scale regularization loss To prevent the degradation of s and b.

9. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements a single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements a single-frame monocular absolute depth estimation method with scale-bias dual-parameter self-calibration as described in any one of claims 1-8.

Images