Data enhancement method for binocular depth estimation data set

Abstract

The invention discloses a target scene depth distribution-oriented binocular depth estimation data set enhancement method, and belongs to the field of computer vision. The method comprises the following steps: analyzing depth probability distribution of a target scene and a public data set; a two-degree-of-freedom image transformation module based on binocular geometry is designed, and depth distribution is systematically changed through horizontal translation and stretching; furthermore, a random sampling strategy of transformation parameters is generated through an optimization method, so that depth distribution of enhanced data approaches target distribution. According to the method, the depth characteristics of the existing data set can be adapted to a specific application scene at low cost on the premise of not collecting new data, and the generalization ability and precision of the binocular depth estimation model in actual deployment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, deep learning and 3D vision technology, and specifically relates to a data augmentation method for binocular stereo matching and depth estimation tasks. It aims to improve the generalization performance of the model in real-world application scenarios through adaptive transformation of depth distribution. Background Technology

[0002] Binocular depth estimation is a core task in computer vision, aiming to calculate the depth information of points in a scene using images captured by a pair of horizontally placed cameras (left and right eyes). This technology is widely used in many fields such as autonomous driving, robot navigation, augmented reality, and 3D reconstruction. With the rapid development of deep learning technology, binocular depth estimation methods based on convolutional neural networks and the Transformer architecture have replaced traditional methods and become the mainstream technology.

[0003] However, the superior performance of deep learning models heavily relies on large-scale, high-quality training datasets. Currently, academia and industry mainly rely on several publicly available benchmark binocular depth estimation datasets for model training and evaluation, but these datasets each have their own characteristics and limitations: the KITTI series datasets for autonomous driving are collected from outdoor road environments, with a depth range concentrated between 0-80 meters, exhibiting a long-tailed distribution, with a predominance of mid-to-long-distance targets and fewer near-distance objects; SceneFlow, as a large-scale synthetic dataset, although it has a wide coverage and high rendering quality, has statistical characteristics of data distribution that differ significantly from the real world, affecting the model's generalization ability; the Middlebury stereo dataset focuses on high-precision indoor scenes with a shallow depth range (usually 0-5 meters), but its data size is small and scene diversity is limited; while the DrivingStereo dataset, which focuses on driving scenarios, includes different weather and lighting conditions, its depth distribution still focuses on a specific range.

[0004] A common problem with these datasets is that their depth distribution range is relatively fixed, and the probability density function of the depth values ​​has a specific shape. For example, the KITTI dataset exhibits a long-tailed depth distribution, with a large number of pixels corresponding to mid- to long-distance distances; however, real-world applications may require models to have higher accuracy in specific depth ranges (such as 0-10 meters for indoor navigation, 5-100 meters for drone aerial photography, and 0.1-3 meters for industrial inspection). This difference in depth distribution leads to a significant performance degradation in models trained on fixed datasets during actual deployment, i.e., the domain adaptation problem.

[0005] Currently, technological exploration to address the domain adaptation problem mainly unfolds along several directions. Traditional data augmentation techniques focus on image-level appearance transformations, such as random adjustments to brightness, contrast, and saturation in the color space, or color jittering; random geometric transformations in space, such as cropping, scaling, rotation, and flipping; or injecting Gaussian noise, salt-and-pepper noise, and blurring. While these methods can improve the robustness of the model to some extent, they struggle to systematically address the core problem of depth distribution mismatch. Although random geometric transformations can alter depth values, because these transformations are unguided and random, they cannot ensure that the transformed depth distribution closely approximates the true distribution characteristics of the target scene.

[0006] Another approach is to employ domain adaptation algorithms. In recent years, researchers have proposed methods including feature-level adaptation (such as using adversarial learning and maximum mean difference techniques in the feature space for domain alignment), output-level adaptation (statistical distribution alignment of the predicted depth map), and style transfer (transferring the appearance style of the target domain to the source domain data). These methods alleviate domain bias to some extent, but they are typically accompanied by complex network designs, instability in the training process, and significant computational overhead. More importantly, most of them focus on bridging appearance style differences, failing to directly and explicitly correct the differences in the depth distribution itself.

[0007] Using computer graphics to generate synthetic data is also a feasible approach, but it has significant limitations: there is an inherent domain gap between synthetic data and real data, and even if the appearance is realistic, its physical properties and noise patterns are different from real data; the generation of high-quality synthetic data depends on professional 3D modeling, rendering technology and a large amount of computing resources; and the depth distribution of synthetic data often still needs to be designed manually, making it difficult to accurately match the distribution characteristics of a specific target scene.

[0008] Furthermore, unsupervised and self-supervised learning based on principles such as photometric consistency reduces the dependence on labeled data. However, these methods tend to experience significant performance degradation in scenes with scarce textures and drastic lighting changes; their training typically relies on unlabeled data of the target scene, making it difficult to fully utilize existing large-scale labeled datasets; and the training process often converges slowly and lacks stability.

[0009] In summary, existing technologies still face the following major shortcomings: First, they cannot systematically and specifically adjust the depth distribution to match the real-world requirements of the target scene; second, the augmentation process often lacks reliable theoretical guidance and physical interpretability; third, many methods are costly and complex to implement; and fourth, the resulting performance improvements are limited and fail to meet the high-precision requirements of practical deployments. Therefore, the industry urgently needs a targeted, interpretable, controllable, and low-cost data augmentation method that can address the depth characteristics of the target scene, thereby fully utilizing existing large-scale labeled datasets and rapidly adapting to various practical application scenarios. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a data augmentation method for binocular depth estimation datasets. This method can systematically transform the depth distribution of publicly available datasets to a range that matches the target application scenario, efficiently generating training data adapted to the depth characteristics of specific scenarios without collecting new data. Simultaneously, this invention provides a physically geometrically based, interpretable depth distribution transformation framework, which can significantly improve the accuracy and robustness of binocular depth estimation models in target scenarios.

[0011] To achieve the above objectives, this invention proposes a stereo depth estimation dataset augmentation method that adapts to the depth distribution of the target scene. The core of the method is to construct a geometric transformation pipeline guided by the depth distribution probability density function, specifically including the following steps: Step S1: Target Domain Depth Distribution Analysis Collect or acquire depth data samples of the target application scenario, perform statistical analysis on the samples, and fit the depth distribution probability density function of the target scenario. .in, This represents the depth value.

[0012] Step S2: Deep Distribution Analysis of Source Dataset The ground truth depth values ​​of the selected public stereo depth estimation dataset (source dataset) are analyzed, and its depth distribution probability density function is obtained by fitting. .

[0013] Step S3: Design and parameterization of depth-guided geometric transformation module Design a two-degree-of-freedom image transformation module based on the geometric principles of binocular vision. This module applies a horizontal translation to the right (or left) image. Horizontal stretching Transformation is used to equivalently and deterministically change the disparity distribution of the entire image pair, thereby achieving scaling and translation of the depth distribution.

[0014] For the right eye image The transformed image for:

[0015] Accordingly, the original disparity map After a completely identical transformation, the new disparity map satisfy:

[0016] According to the stereo depth calculation formula ,in Baseline length Focal length For parallax, the new depth value can be determined. Compared with the original depth value The relationship is:

[0017] when When transformed into pure scale scaling, ;when When the transformation is a pure translation, and This is a non-linear relationship. It is achieved through joint adjustment. It can perform complex nonlinear mappings on depth distributions.

[0018] Step S4: Generation of random transformation strategy for distribution matching The target distribution obtained based on steps S1 and S2 Source distribution Construct an optimization problem to find transformation parameters. The sampling strategy makes the depth distribution of the transformed source data... Get as close as possible to the target distribution.

[0019] Define distribution difference measure For example, using the one-dimensional Wasserstein distance:

[0020] By minimizing It can solve for an optimal static parameter pair for the entire dataset. .

[0021] To enhance data diversity, a better strategy is to learn a parameter distribution. (e.g., a two-dimensional Gaussian mixture model), its parameters The following optimization objectives were learned:

[0022] in, This is a regularization term used to constrain the rationality of transformations (such as avoiding excessive image distortion). This is the regularization coefficient.

[0023] During the training process, the learned distribution Data augmentation is performed using randomly sampled parameters.

[0024] Step S5: Enhanced Data Generation and Model Training In each iteration of model training, a batch of raw stereo image pairs is first loaded. and its parallax true value Next, the policy distribution obtained from step S4... Mid-sample a set of transform parameters Subsequently, formula (1) was used to process the right eye image. Perform geometric transformations to obtain The same parameters must be used throughout this process. The same transformation method is used for the parallax truth. Perform a synchronous transformation to generate a new monitoring signal. As shown in Equation (2). Finally, based on the enhanced training samples... Calculate the loss function and update the parameters of the stereo depth estimation model.

[0025] Compared with the prior art, the present invention has the following significant advantages: First, in terms of accurate adaptation of depth distribution, the present invention is the first to start from the probability density function level, guide data augmentation by solving the distribution matching problem, realize the systematic and targeted transfer of depth distribution from the source domain to the target domain, thereby solving the fundamental problem of insufficient model generalization ability caused by depth distribution mismatch.

[0026] Secondly, this invention is based on physical principles and has high interpretability. The geometric transformation used strictly follows the binocular stereo vision model. The transformed image pairs still satisfy the epipolar geometric constraints, and the generated depth information is physically coherent and reasonable, which enhances the credibility of the method.

[0027] Third, this method is low-cost and easy to implement, requiring only a small number of depth samples from the target scene for distribution analysis. This avoids expensive and time-consuming large-scale intensive data collection and fine annotation, significantly reducing the cost of domain adaptation. Furthermore, this invention is plug-and-play and highly versatile, capable of serving as an independent data preprocessing or online enhancement module, seamlessly integrating with various deep learning-based binocular depth estimation models without relying on specific network structures.

[0028] Fourth, and finally, this method can bring significant performance improvement. By "calibrating" the depth distribution of the training data to the target scene, it effectively narrows the gap between the training domain and the test domain, thereby achieving better depth estimation accuracy and robustness in the target scene than traditional augmentation methods or direct transfer methods. Attached Figure Description

[0029] Figure 1 The overall flowchart of this invention includes steps such as target scene and source data depth distribution analysis, geometric transformation module design, distribution matching strategy generation, augmented data generation and model training.

[0030] Figure 2 : A schematic diagram of depth distribution transformation, illustrating the mathematical process of approximating the target distribution from the source distribution through geometric transformation.

[0031] Figure 3 : A block diagram illustrating the principle of depth-guided geometric transformation, showing how horizontal translation and stretching can change parallax and depth distribution.

[0032] Figure 4 : Comparison diagram of the depth estimation process of the present invention and the baseline method in the target scene. Detailed Implementation

[0033] The present invention will now be described in detail with reference to a preferred embodiment and accompanying drawings. This embodiment takes adapting an autonomous driving scenario (KITTI dataset) to an indoor navigation scenario as an example to demonstrate the complete implementation process of the present invention. The experiment was conducted on a workstation equipped with an NVIDIA RTX 4090 GPU, an Intel Core i9-13900K processor, 64GB DDR5 memory, and 2TB NVMe SSD storage. The software environment was based on the Ubuntu 20.04 LTS operating system, using the PyTorch 1.12.1 deep learning framework with the CUDA 11.6 acceleration library, and scientific computing libraries such as NumPy, SciPy, and OpenCV. Regarding the dataset, the source data used the KITTI 2015 stereo matching dataset, containing 400 image pairs from 200 training scenarios and dense disparity ground truth values; the target scene data was selected from the NYU Depth V2 dataset, consisting of 1000 indoor scene RGB-D images as statistical samples. Figure 1 This is the overall flowchart of the present invention.

[0034] Step S1: Target Domain Depth Distribution Analysis (Detailed Implementation) During the data preprocessing stage, 1000 depth maps were randomly selected from the NYU Depth V2 dataset. First, each depth map underwent a validity check, and invalid depth values ​​were removed. Since this dataset was acquired using a Kinect sensor, its depth range is limited by hardware to between 0.5 and 10 meters; therefore, sensor calibration parameter correction was necessary.

[0035] Depth values ​​were then extracted and statistically analyzed. All valid depth pixel values ​​were flattened into a one-dimensional array, yielding approximately 30 million depth sample points. Basic statistics were calculated: a minimum value of 0.48 meters, a maximum value of 9.87 meters, a mean of 2.34 meters, and a standard deviation of 1.56 meters. A depth histogram was plotted with a group interval of 0.1 meters, allowing for a visual observation of the distribution pattern.

[0036] In the probability density function estimation stage, a kernel density estimation method is used to fit the continuous probability density function. Specifically, a Gaussian kernel function is chosen, and the bandwidth is automatically determined using the Silverman rule. The bandwidth calculation formula defined by this rule is as follows: ,in The standard deviation of the sample is 1. The sample size is given. During implementation, to improve numerical stability, the depth samples are first transformed to logarithmic space. Then, a Gaussian kernel density estimator is used to fit the transformed data.

[0037] Finally, a probability density function lookup method is defined. For any depth value, it is transformed to logarithmic space, and then the probability density is calculated using a fitted kernel density estimation model. Finally, it is transformed back to the original depth space. The fitted target domain depth probability density function is then obtained. The data shows that the depth distribution is concentrated in the range of 1-4 meters, with a peak at about 1.8 meters, and exhibits a right-skewed distribution, which is consistent with the actual situation in indoor scenes where most objects are located at a medium distance.

[0038] Step S2: Deep Distribution Analysis of Source Dataset (Detailed Implementation) First, the KITTI data is processed: all disparity ground truth files from the KITTI 2015 training set are loaded, and using the calibration parameters provided by KITTI (baseline length 0.54 meters, focal length 721.5 pixels), the disparity values ​​are converted into depth values ​​using the following formula: ,in For depth value, This represents the disparity value. Subsequently, all invalid disparity points with a disparity value of zero are removed.

[0039] Next, a depth statistical analysis was conducted: approximately 400 million valid depth samples were extracted from all 200 training scenes, and the minimum depth was 0.85 meters, the maximum depth was 79.3 meters, the mean depth was 28.7 meters, and the standard deviation was 22.4 meters. The analysis revealed that the KITTI depth distribution exhibits a clear long-tail characteristic, with a large number of pixels corresponding to medium to long distances of 20 to 60 meters.

[0040] In the source distribution modeling stage, the kernel density estimation method is also used to fit the source domain depth distribution. Because the KITTI depth range is quite large, a logarithmic spatial transformation is performed on the depth values ​​to improve the fitting effect. In practice, the KITTI depth values ​​are first logarithmically transformed, then an adaptive bandwidth selection method is used to determine the bandwidth parameter, and finally a Gaussian kernel density estimator is used to fit the logarithmically transformed depth data.

[0041] Step S3: Implementation details of the geometric transformation module First, determine the range of transformation parameters. Based on physical constraints and image quality considerations, set the horizontal stretching coefficient. The value ranges from 0.7 to 1.5, representing the horizontal translation amount. The value range is from -30 to 30 pixels. This parameter range can cover a sufficient depth transformation space while avoiding severe image distortion.

[0042] In the specific implementation of image transformation, a normalized sampling grid is first created, where... The coordinate range is from -1 to 1, corresponding to the image width. The coordinates range from -1 to 1, corresponding to the image height. When applying the transformation formula, for each output pixel position... The corresponding input pixel position is ,in The image width is [value]. Bilinear interpolation is used to resample the right eye image to ensure smoothness and avoid jagged artifacts after transformation. The same spatial transformation is applied to the disparity map, but nearest-neighbor interpolation is used to preserve the discrete nature of the disparity values. After transformation, the disparity values ​​are calculated according to the formula [formula missing]. Update, in which For the original parallax, For new parallax.

[0043] To ensure the correctness of the transformation process, the depth transformation relationship was verified through theoretical derivation to ensure that the new depth value and the original depth value conform to the relationship defined by formula (3). At the same time, various test cases were designed to verify the stability of boundary conditions, including transformation tests on minimum and maximum depth values, and a systematic evaluation of the transformation effect under parameter boundary conditions.

[0044] Step S4: Generation of Distributed Matching Strategy (Detailed Optimization Process) To effectively measure the difference in depth distribution between the source and target domains, a suitable distribution distance metric must be selected. After comparing various metrics such as KL divergence, JS divergence, and Wasserstein distance, this method ultimately chooses the one-dimensional Wasserstein distance. This distance not only effectively handles the case where the distributed support sets do not overlap, but also possesses good mathematical properties and robustness, making it more suitable for the requirements of this task.

[0045] In the process of parameter space discretization and optimization, the parameter space is first discretized into a grid: the horizontal stretching coefficient is then... Take 21 equally spaced points within the interval 0.7 to 1.5, and calculate the horizontal translation amount. Sixty-one equally spaced points were selected within the range of -30 to 30 pixels, generating a total of 1281 parameter combinations. For each parameter combination... The theoretical transformed depth distribution is calculated. Specifically, the transformation evaluation is performed on the depth value grid, the transformed depth value is calculated according to formula (3), the probability density transformation is completed through Jacobian determinant, and finally the transformed probability density is normalized.

[0046] The Wasserstein distance is calculated using the corresponding functions in the SciPy library. For each combination of parameters, a transformed empirical distribution and a target empirical distribution are constructed, and then the Wasserstein distance between the two distributions is calculated. This distance reflects the minimum "work" required to transform the source distribution into the target distribution.

[0047] The optimization strategy is generated by traversing all parameter combinations and calculating the Wasserstein distance for each combination. The top 20 parameter combinations with the smallest distances are selected as a candidate set, and a hybrid sampling strategy is constructed. This strategy consists of three parts: sampling from the candidate set with an 80% probability based on the inverse distance weighted average (i.e., combinations with smaller distances have a higher sampling probability); sampling with a 15% probability around the candidate parameters using Gaussian perturbation; and sampling with a 15% probability around the candidate parameters. and The perturbation standard deviations are set to 0.05 and 2 pixels, respectively; uniform sampling is performed throughout the parameter space with a probability of 5% to maintain a certain exploration capability and avoid getting trapped in local optima.

[0048] Finally, the generation strategy was validated. 1000 sets of parameters were sampled from the strategy, and their corresponding average Wasserstein distances were calculated. The transformed depth distribution was then compared with the target distribution using visualization. By plotting the distribution curves, the degree of matching between the transformed and target distributions can be intuitively evaluated, ensuring that the distribution matching effect meets the requirements.

[0049] Step S5: End-to-end training and integration In terms of model selection and configuration, this embodiment uses PSMNet (Pyramid Stereo Matching Network) as the baseline model. The model is configured as a stack of three hourglass networks, with a feature dimension of 32. The loss function is a smoothed L1 loss, which is optimized by combining it with disparity regression loss.

[0050] During the training pipeline integration process, a deep distribution adaptive augmentation module was constructed, which includes policy parameters and corresponding sampling methods. In each batch loop of training, transformation parameters are first sampled from a predefined policy. Subsequently, the geometric transformation described in step S3 is applied to each sample within the batch to obtain the enhanced right eye image and disparity ground truth. The sampling method strictly follows the mixing strategy defined in step S4: a random number is generated; if it is less than 0.8, samples are taken from the candidate set using distance-weighted sampling; if it is between 0.8 and 0.95, Gaussian perturbation sampling is performed around the candidate parameters; otherwise, uniform sampling is performed across the entire parameter space. The enhanced data batch will be used for model training.

[0051] The training hyperparameters were set as follows: batch size of 8 (limited by GPU memory), initial learning rate of 0.001, and the optimizer used was Adam (where... , The learning rate scheduling strategy is to reduce the learning rate to 0.7 times its original value after every 20 training rounds, with a total of 80 training rounds.

[0052] After completing end-to-end training, we conducted a systematic experimental evaluation and results analysis of the model. First, we established an evaluation benchmark. The target scene test set was selected from 654 images in the NYU Depth V2 dataset. The evaluation metrics included RMSE (root mean square error) to measure the absolute accuracy of depth estimation, Abs Rel (absolute relative error) to reflect relative accuracy, and δ1 (the percentage of pixels whose predicted depth is within 1.25 of the ground truth) to represent the accuracy of depth estimation.

[0053] We set up four sets of comparative experiments: baseline 1 is a PSMNet model trained using the original KITTI data; baseline 2 is a PSMNet trained using traditional data augmentation (including color jitter and random cropping); baseline 3 is a PSMNet trained using the domain adaptation method ADAS; and the present invention is trained using the proposed augmentation method.

[0054] Quantitative results show that the present invention achieves the best performance on the NYU Depth V2 test set, with an RMSE of 0.621 m, an Abs Rel of 0.158, and a δ1 of 75.2%, which is comprehensively superior to all baseline methods.

[0055] Table 1: Performance Comparison on the NYU Depth V2 Test Set

[0056] Qualitative analysis shows that the present invention exhibits more accurate depth estimation capabilities in indoor scenes: nearby objects have richer details and clearer edges; depth boundaries are sharper, reducing blurring; and depth estimation in large planar areas is smoother and more consistent, effectively suppressing noise and artifacts. In contrast, baseline methods are prone to artifacts at abrupt depth changes, while the present invention, due to its more consistent depth distribution, performs more stably in these areas.

[0057] The effectiveness of the strategy components was further verified through ablation experiments: the RMSE was 0.658 m when only the optimal static parameters were used; the RMSE was 0.635 m when only the candidate set sampling was used; and the full hybrid strategy achieved the best results with an RMSE of 0.621 m, proving the rationality of the hybrid strategy design.

[0058]

[0059] Computational efficiency analysis shows that the augmentation operations are performed in real time on the GPU, adding only about 3 milliseconds of processing time per batch; in terms of training efficiency, the present invention has achieved performance superior to the baseline method after 80 training epochs in 40 training epochs; in terms of memory overhead, it only adds negligible computational overhead for transform parameter sampling and image resampling.

[0060] In addition to the above-described embodiments, this invention can be extended to various implementation variations to meet different application requirements. In multi-target scene adaptation scenarios, when the target environment contains multiple depth distribution patterns—for example, when a UAV needs to handle both near-ground and distant targets simultaneously—a hybrid distribution modeling method can be used to represent the target distribution as follows: ,in The mixed weights for each sub-distribution, For the first There are several sub-distributions. Based on this, for each sub-distribution... Learn the corresponding transformation strategies respectively And in the data augmentation stage, with probability By selecting appropriate sampling strategies, collaborative adaptation of multimodal depth distribution can be achieved.

[0061] For application scenarios where data is gradually accumulated, an incremental adaptation strategy can be adopted. In the initial stage, a small number of target samples are used to estimate the depth distribution. Subsequently, as new data is continuously acquired, the target distribution estimate is incrementally updated online, and the transformation strategy is dynamically adjusted to achieve a gradual and adaptive distribution matching process.

[0062] Furthermore, this invention can be organically combined with existing data augmentation technologies to construct a multi-layered enhancement process. Specifically, this invention can be applied first for depth distribution adaptation, followed by traditional image enhancement operations such as color dithering and random cropping, thereby simultaneously addressing depth distribution differences and appearance style differences at the system level, forming a more comprehensive data augmentation solution.

Claims

1. A data augmentation method for a stereo depth estimation dataset, characterized in that, Includes the following steps: S1: Analyze the depth data of the target application scenario to obtain the depth distribution probability density function of the target scenario. ; S2: Analyze the depth data of the publicly available stereo dataset to obtain the depth distribution probability density function of the source dataset. ; S3: Construct a depth-guided geometric transformation module that applies a horizontal translation to one image in a stereo image pair. and horizontal stretching The similarity transformation is used to change the depth distribution corresponding to the image pair; S4: Based on the above and A random sampling strategy is generated for the parameters of the geometric transformation module, such that the depth distribution of the source data after transformation by the module is optimized. With the Minimize the differences between them; S5: During model training, transformation parameters are obtained according to the random sampling strategy, the original training samples are augmented online using the geometric transformation module, and the augmented samples are used to train the stereo depth estimation model.

2. The data augmentation method for a stereo depth estimation dataset according to claim 1, characterized in that: In step S3, the geometric transformation is applied to the right eye image. The transformation formula is: Simultaneously, for the corresponding disparity ground truth map Perform synchronous transformation: .

3. The data augmentation method for a stereo depth estimation dataset according to claim 1 or 2, characterized in that: In step S4, the difference minimization is achieved by optimizing a distribution distance metric. To achieve this, the distribution distance metric includes KL divergence, Wasserstein distance, or histogram intersection.

4. The data augmentation method for a stereo depth estimation dataset according to claim 3, characterized in that: In step S4, the random sampling strategy is obtained by solving the following optimization problem: learning a parameter distribution. This causes the expected distribution difference resulting from the parameters sampled from this distribution. Minimum.

5. The data augmentation method for a stereo depth estimation dataset according to claim 1, characterized in that: In step S1, the depth distribution probability density function of the target scene is obtained by kernel density estimation or Gaussian mixture model fitting of its small-scale depth samples.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 5.

Images