CN114550491A - Airport underground parking space positioning and navigation method based on mobile phone - Google Patents

Abstract

A mobile phone-based method for locating and navigating underground parking spaces in an airport can use the mobile phone's own vision and inertial sensors to perform parking space locating and navigation algorithms. The present invention firstly locates the parking space photographed by the mobile phone camera through the visual positioning method based on scene coordinate regression, the visual positioning algorithm realizes the input RGB image and outputs the camera pose by the end-to-end neural network algorithm, and then when personnel need to find the parking space, Take the current position image again to obtain the current initial position of the person, and then use the inertial sensor to estimate the position in real time for navigation, and use the camera positioning pose as an opportunity signal to compensate the accumulated error of the inertial sensor, and finally guide the person to the parking space. The present invention realizes real-time high-precision positioning and navigation through fusion positioning of high-precision visual positioning based on scene coordinate regression and pose estimation of high-frequency inertial sensors.

Description

A mobile phone-based method for locating and navigating underground parking spaces in airports

technical field

The invention belongs to the field of indoor positioning, and in particular relates to a mobile phone-based method for positioning and navigating underground parking spaces in airports.

Background technique

In recent years, with the rapid development of the economy, the number of cars has greatly increased, and the use of cars has become more and more popular. Parking has become a rigid need for car travel. In order to cope with the surge in demand for parking spaces, underground parking comes along with it. With the continuous construction and increase of the parking lot, and its scale is also getting larger and larger, when the car owner performs other activities after parking, it becomes very difficult to return to the parking lot to find the vehicle, so the positioning and navigation service for the large-scale parking lot It is essential, especially in the underground parking lot of the airport. It is generally used to drive a private car to park and then fly to the field and then return to the parking lot to find a vehicle. After a long time, it becomes more difficult to find a parking location.

At present, there are many platforms mainly developed for outdoor positioning and navigation technology on the market, and most of them are very mature. Generally, the navigation and positioning of the parking lot on the ground mainly adopts the GPS positioning system, and the system will provide the vehicle owner with the position information of the vehicle. In addition, the location information of the car owner can also be obtained, and the position information of the car owner obtained by GPS positioning can be used to plan the best navigation scheme to reach the parking space according to the relative position between the parking space and the car owner, and search for navigation and positioning. vehicle. However, in the underground parking lot, the satellite signal often becomes very weak, so the positioning technology is generally ineffective or even unavailable. At present, the indoor navigation technology for the underground parking lot in my country is relatively rare, and most of the existing indoor positioning and navigation technologies are equipped with equipment. It is expensive or has low positioning accuracy, so it is very necessary to use the indoor positioning technology of mobile phones to apply the parking space navigation and positioning technology of the airport underground parking lot. The positioning and navigation based on the popular mobile phone terminal is popular and feasible.

SUMMARY OF THE INVENTION

In view of the above problems, the present invention aims to provide a mobile phone-based method for locating and navigating an airport underground parking space, which is used to realize indoor positioning and navigation technology in an underground parking lot, and can meet the low-cost and accurate navigation performance of a common mobile phone.

The invention provides a mobile phone-based method for locating and navigating underground parking spaces in an airport, the specific steps are as follows, and it is characterized in that

The image acquisition unit of the S11 mobile phone collects images, and performs coordinate regression on the acquired RGB image through the differentiable RANSAC method to obtain the pose prediction of the center point of the image.

The differentiable RANSAC methods described above include:

S11-1 scene coordinate regression, predict the scene coordinate y _i (ω) of the 3-dimensional parking lot corresponding to each pixel i of the 2-dimensional RGB image through the ResNet network, where ω is the parameter of the neural network ResNet model, through which the image is realized 2-dimensional pixel point (x _i , y _i ) to (R, T) of 3-dimensional scene coordinates, R represents the spatial rotation of 3 degrees of freedom, T represents the displacement of 3 degrees of freedom, which is represented by f(ω) The mapping from the 2D image camera coordinate system to the 6D pose of the camera, where ω is the parameter that the ResNet neural network needs to learn, and the learnable parameter ω is optimized by minimizing the final estimated expected pose loss l on the training set:

In the formula, f ^* represents the true value of the pose of the image I. In order to train the neural network through the gradient descent method to optimize the parameter ω, the derivative of the parameter ω is obtained. The partial derivative of the above formula is:

S11-2 pose hypothesis sampling, the generation of each hypothesis f(ω) is generated by the subset corresponding to the image, the size of this subset is the minimum number of mappings required to calculate a unique solution, and the minimum subset is M _J , where J={j ₁ ,L,j _n }, where n is the minimum number of subsets; suppose the scene regression coordinates of f(ω) are PnP problems, so four scene coordinates are enough to define a unique camera pose, so n=4, since the result of randomly selecting four points for prediction each time may be wrong, by randomly selecting 4 pairs of images to the scene coordinates for prediction, the set of pose prediction f(ω) of n images can finally be generated , each generated scene coordinate hypothesis f(ω) depends on the parameter ω;

S11-3 Select the optimal pose hypothesis. For the hypothesis f(ω) with different scene coordinate regression performance, an evaluation mechanism is required to select the optimal hypothesis. The evaluation determines the selection of the camera pose hypothesis and the effect of the final refinement of the hypothesis. to produce the final estimate, as follows;

First define the pixel i of the image and assume the reprojection error of f(ω) as:

e _i (f,ω)=||Cf ^-1 k _i (ω)-p _i ||

where pi is the image coordinate (x _i , y _i ) of each pixel _i of the image, C is the camera projection matrix, _ki is the interior point, if e _i <τ, where τ determines the error threshold of the interior point;

The scene coordinate regression work relies on counting the inliers k _i to score the hypothesis. To achieve end-to-end training of the neural network, a differentiable function is constructed by using the sigmoid function:

where e _i represents the reprojection error, and the hyperparameter β controls the softness of the sigmoid;

The evaluation function p(f) scores the consistency of each hypothesis with all scene coordinate predictions, and the hypothesis f _j (ω) with index j is selected according to the probability distribution P(j, ω, α) derived from the score value, Hypotheses with high scores are more likely to be selected, and the final hypothesis is selected according to the softmax distribution P:

The size of the inlier count scores can vary greatly depending on the difficulty of the scene, usually by a large order of magnitude in different environments, keeping the evaluation scores within a reasonable range for having a wide distribution P(j;ω, α) is important, important for stable end-to-end training, manually setting the hyperparameter α for each scene is a tedious task, automatically adapting α during end-to-end training, and choosing the size of the hyperparameter α by entropy :

In end-to-end training, the training is done by argmin _α |S(α)-S ^* | gradient descent to determine the parameter α, the choice is to build the target entropy in the first iteration of end-to-end training and keep it stable throughout the process ;

S11-4 refines the optimal pose hypothesis. The refinement function R is an iterative process that alternates between determining inlier pixels using the current pose estimate and optimizing the estimation of inlier pixels. If a neural network is to be used in the iterative process, End-to-end training must be realized. In order to improve the generalization and accuracy of the model obtained from training, the training process is divided into two steps. The first step uses the MAML method to divide the training process into an inner loop and an outer loop, so that the model has a better performance. Initial parameters, the second part is differentiable for the refinement process, and finally the neural network model can output the final pose R(y _j (ω));

S21 obtains the pose of the current person by performing scene coordinate regression on the image captured by the mobile phone. The coordinate regression method of the differentiable RANSAC has very high accuracy, and the pose accuracy error can be within 5cm and 5°. The position and attitude are used as the current initial position.

S22 obtains the data of the built-in IMU of the mobile phone at this moment, performs inference and update according to the obtained initial position and attitude data, that is, the attitude data of the previous moment, and obtains the pose data of the current moment,

S23 uses the IMU to update the pose and navigate the process, and uses the pose obtained by the scene coordinate regression of the image as the new initial coordinate to initialize the iterative update process of the IMU, and eliminate the accumulated error of the inertial sensor through high-precision visual positioning.

As a further improvement of the present invention, the first step of the training process in step S11-4 adopts the MAML method to divide the training process into an inner loop and an outer loop as follows:

For model initialization, MAML is introduced to initialize the model parameters and make the model have a better initial performance. The training process of the neural network is divided into an inner loop and an outer loop. The inner loop realizes the training of basic model functions, and the outer loop improves the generalization of the model. training, and also divide the training set into two parts for the training of two loops;

The training set of the inner loop is initialized with the real value of the rendered scene when there is a 3D model, otherwise the approximated real value is initialized heuristically. The optimization of the inner loop is:

The specific parameter iteration process of the inner loop is expressed as:

ω'=ω-μ▽ _ω L(f _ω )

In the formula, ω' is the optimal parameter of the inner loop iteration, ω is the initialization parameter, μ is the learning rate of the inner training process, and ▽ _ω L(f _ω ) represents the gradient of the inner loop.

For the outer loop, using the reprojection error calculated by the ground truth pose, and using a heuristic algorithm in the inner loop step, effectively restores the correct depth of the scene points;

Therefore, the optimization process of the inner loop is:

Regarding the transfer of parameters of the inner loop and the outer loop, the specific parameter transfer method is to perform meta-update or meta-optimization before sampling the next batch of tasks, find the optimal parameter ω through the training of the inner loop, and then calculate the optimal parameter ω for each inner loop. The gradient of ω, and the random initialization parameter ω is updated by the gradient update method, which makes the random initialization parameter θ find an initial parameter that is more similar to the target task, and does not need to take many gradient steps when training the next batch of tasks;

The whole process is described as follows:

是对于外循环对于参数ω'的梯度；where ω is the initialization parameter, η is the learning rate hyperparameter of the outer loop,

is the gradient of the outer loop for the parameter ω';

For the second step of training, training the model in an end-to-end manner requires all processes to be differentiable, including pose optimization. For the process of refining the hypothesis, the process of refining the pose is:

Since the argmin function is not differentiable, the Gauss-Newton method is used to obtain the differentiation of the refinement function:

其中h_o＝R^t＝∞(y)，最终实现端到端的训练。In the formula, J _e represents the Jacobian matrix, which contains the partial derivatives

where h _o = R ^{t = ∞} (y), and finally achieve end-to-end training.

As a further improvement of the present invention, each hypothesis y(ω) depends on the parameters of the corresponding scene coordinates.

As a further improvement of the present invention, the hyperparameter α controls the width of the distribution, and selecting an appropriate hyperparameter is helpful for end-to-end learning, and its function is to reduce the difference between the hypothesis scores, so that the gradient values are not too far apart .

Aiming at the problems of low precision or high cost of the current positioning and navigation system in the indoor environment of the underground parking lot, the present invention proposes a mobile phone-based positioning and navigation method for the underground parking space of the airport. The application of this method is implemented in two steps. The first step is to locate the image captured by the mobile phone through the visual positioning method of differentiable RANSAC when the person is parking, and the parking position of the vehicle is obtained as the end coordinate of the subsequent navigation and positioning; When people enter the underground parking lot to find vehicles again, they first obtain the initial coordinates of the navigation through the visual positioning of the images captured by the mobile phone, which is also used as the initial coordinates of the mobile phone inertial sensor IMU.

For the indoor positioning, navigation and positioning problem, the present invention proposes a method of fusion of differentiable RANSAC visual positioning and inertial positioning, which can integrate high-precision visual positioning and high-frequency inertial positioning, effectively solve the problem of low positioning accuracy without GPS in indoor positioning, and at the same time can The realization of navigation and positioning on the mobile phone platform can be widely used in the parking space navigation service of the airport underground parking lot.

Description of drawings

1 is a general flow chart of a mobile phone-based airport underground parking space positioning and navigation method of the present invention;

Figure 2 is a schematic diagram of a differentiable RANSAC scene regression visual localization algorithm.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments:

A mobile phone-based method for locating and navigating underground parking spaces in airports of the present invention, as shown in FIG. 1 , the present invention is an indoor positioning method integrating vision and inertia, including the following steps:

The image acquisition unit of the S11 mobile phone collects images, and performs coordinate regression on the acquired RGB image through the differentiable RANSAC method to obtain the pose prediction of the center point of the image.

Figure 2 shows the differentiable RANSAC method, and its specific implementation method includes:

S11-1 scene coordinate regression, predict the scene coordinate y _i (ω) of the 3-dimensional parking lot corresponding to each pixel i of the 2-dimensional RGB image through the ResNet network, where ω is the parameter of the neural network ResNet model, the main function of the model It is to realize the (R, T) of the 2-dimensional pixel point (x _i , y _i ) of the image to the 3-dimensional scene coordinate, R represents the spatial rotation of 3 degrees of freedom, and T represents the displacement of 3 degrees of freedom, which is represented by f( ω) represents the mapping from the 2-dimensional image camera coordinate system to the 6-dimensional pose of the camera, where ω is the parameter that the ResNet neural network needs to learn, and the learnable parameter ω is optimized by minimizing the final estimated expected pose loss l on the training set:

In the formula, f ^* represents the true value of the pose of the image I. In order to train the neural network to optimize the parameter ω through the gradient descent method, we need to derive the parameter ω. The partial derivative of the above formula is:

S11-2 pose hypothesis sampling, the generation of each hypothesis f(ω) is generated by the subset corresponding to the image, the size of this subset is the minimum number of mappings required to calculate a unique solution, and the minimum subset is M _J , where J={j ₁ , L, j _n }, where n is the minimum number of subsets. This paper assumes that the scene regression coordinates of f(ω) are PnP problems, so four scene coordinates are enough to define a unique camera pose, so n=4. However, the result of randomly selecting four points for prediction each time may be wrong, so by randomly selecting 4 pairs of images to the scene coordinates for prediction, a set of pose prediction f(ω) of n images can finally be generated, each generating The scene coordinates of f(ω) are assumed to depend on the parameter ω.

S11-3 Select the optimal pose hypothesis. For the hypothesis f(ω) with different scene coordinate regression performance, an evaluation mechanism is required to select the optimal hypothesis. The evaluation determines the selection of the camera pose hypothesis and the effect of the final refinement of the hypothesis. to produce the final estimate. First of all, the present invention defines the reprojection error of the pixel i of the image and the assumption f(ω) as:

e _i (f,ω)=||Cf ^-1 k _i (ω)-p _i ||

where pi is the image coordinate (x _i , y _i ) of each pixel _i of the image, C is the camera projection matrix, _ki is the interior point, if e _i <τ, where τ determines the error threshold of the interior point.

The scene coordinate regression work of the present invention relies on the counting of the interior points k _i to score the hypothesis. In order to realize the end-to-end training of the neural network, we construct a differentiable function by using the sigmoid function:

where e _i represents the reprojection error, and the hyperparameter β controls the softness of the sigmoid.

The evaluation function p(f) scores each hypothesis for its agreement with all scene coordinate predictions. The hypothesis f _j (ω) with index j is selected according to the probability distribution P(j,ω,α) derived from the score value. Hypotheses with high scores are more likely to be selected, and the final hypothesis is selected according to the softmax distribution P:

The size of the inlier count scores can vary widely depending on the difficulty of the scene, often by a large order of magnitude across different environments. Keeping the evaluation scores within a reasonable range is important for having a wide distribution P(j;ω,α), which is important for stable end-to-end training. Manually setting the hyperparameter α for each scene is a tedious task, so we automatically adapt α in end-to-end training, using entropy to choose the size of the hyperparameter α:

In end-to-end training, the training is done by argmin _α |S(α)-S ^* | gradient descent to determine the parameter α, the choice is to build the target entropy in the first iteration of end-to-end training and keep it stable throughout the process .

S11-4 refines the optimal pose hypothesis. The refinement function R is an iterative process that alternates between determining inlier pixels using the current pose estimate and optimizing the estimation of inlier pixels. If a neural network is to be used in the iterative process, End-to-end training must be realized. In order to improve the generalization and accuracy of the model obtained from training, the training process is divided into two steps. The first step uses the MAML method to divide the training process into an inner loop and an outer loop, so that the model has a better performance. The initial parameters, the second part is differentiable for the refinement process, and finally the neural network model can output the final pose R(h _j (ω)).

The neural network of the present invention is trained in an end-to-end manner using paired RGB images and ground truth poses, but doing so from scratch will fail because the system quickly reaches local minima. We propose a new two-step training mode, each with a different objective function. Depending on whether a 3D scene model is available, we initialize the network with rendered or approximated scene coordinates in the first step and improve system accuracy. The second step of training improves the accuracy of the system, which is critical when no 3D model is provided for initialization.

The first step of the training process is to introduce MAML to initialize the model parameters and make the model have a good initial performance. The training process of the neural network is divided into an inner loop and an outer loop. The inner loop realizes the training of basic model functions. The training of the outer loop to improve the generalization of the model also divides the training set into two parts for the training of two loops.

The training set of the inner loop is initialized with the real value of the rendered scene when there is a 3D model, otherwise the approximate truth value is initialized heuristically. The optimization of the inner loop is:

The specific parameter iteration process of the inner loop can be expressed as:

ω'=ω-μ▽ _ω L(f _ω )

In the formula, ω' is the optimal parameter of the inner loop iteration, ω is the initialization parameter, μ is the learning rate of the inner training process, and ▽ _ω L(f _ω ) represents the gradient of the inner loop.

For the outer loop, we optimize the reprojection error computed using the ground truth pose. If we use a heuristic in the inner loop step, we can efficiently recover the correct depth of scene points. Therefore, the optimization process of the inner loop is:

Regarding the transfer of parameters in the inner loop and the outer loop, the specific parameter transfer method is that we perform meta-update or meta-optimization before sampling the next batch of tasks. That is to say, in the previous step, the optimal parameter ω is found by training the inner loop, then the gradient of ω for each inner loop is calculated, and our random initialization parameter ω is updated by the gradient update method. This makes it possible to randomly initialize the parameters θ to find an initial parameter that is more similar to the target task, without taking many gradient steps when training the next batch of tasks. The whole process can be expressed as follows:

是对于外循环对于参数ω'的梯度。where ω is the initialization parameter, η is the learning rate hyperparameter of the outer loop,

is the gradient of the outer loop with respect to the parameter ω'.

In the second training step, we want to train our model in an end-to-end fashion. This requires all processes to be differentiable, including pose optimization. For the process of refining the hypothesis, our process of refining the pose is:

Since the argmin function is not differentiable, the Gauss-Newton method is used to obtain the differentiation of the refinement function:

其中h_o＝R^t＝∞(y)，最终我们可以实现端到端的训练。In the formula, J _e represents the Jacobian matrix, which contains the partial derivatives

where h _o = R ^{t = ∞} (y), and finally we can achieve end-to-end training.

S21 obtains the pose of the current person by performing scene coordinate regression on the image captured by the mobile phone. The coordinate regression method of the differentiable RANSAC has very high accuracy, and the pose accuracy error can be within 5cm and 5°. The position and attitude are used as the current initial position.

S22 obtains the data of the built-in IMU of the mobile phone at this moment, performs inference and update according to the obtained initial position and attitude data, that is, the attitude data of the previous moment, and obtains the pose data of the current moment,

S23 uses the IMU to update the pose and navigate the process, and uses the pose obtained by the scene coordinate regression of the image as the new initial coordinate to initialize the iterative update process of the IMU, and eliminate the accumulated error of the inertial sensor through high-precision visual positioning.

The above are only examples of preferred embodiments of the present invention, and are not intended to limit the present invention in any other form, but any modifications or equivalent changes made according to the technical essence of the present invention still belong to the claimed protection of the present invention range.

Claims (4)

1. A mobile phone-based airport underground parking space positioning and navigation method is characterized by comprising the following specific steps

S11, collecting the image by the image collecting unit of the mobile phone end, performing coordinate regression on the obtained RGB image by a differentiable RANSAC method to obtain the position and pose prediction of the central point of the image,

the above differentiable RANSAC method includes:

s11-1 scene coordinate regression, and predicting scene coordinate y of 3-dimensional parking lot corresponding to each pixel i of two-dimensional RGB image through ResNet network_i(omega), wherein omega is a parameter of a neural network ResNet model, and 2-dimensional pixel points (x) of the image are realized through the model_i,y_i) (R, T) to 3-dimensional scene coordinates, R representing spatial rotation of 3 degrees of freedom, T representing displacement of 3 degrees of freedom, wherein the mapping of the 2-dimensional image camera coordinate system to the 6-dimensional pose of the camera is represented by f (ω), where ω is a parameter that the ResNet neural network needs to learn, and the learnable parameter ω is optimized by minimizing the final estimated expected pose loss l on the training set:

in the formula f^*Representing the true value of the pose of the image I, and solving the derivative of the parameter omega in order to train and optimize the parameter omega for the neural network by a gradient descent method, wherein the partial derivative of the formula is as follows:

s11-2 pose hypothesis sampling, each hypothesis f (ω) is generated from a corresponding subset of the image, the subset having a size that is the minimum number of mappings required to compute a unique solution, the minimum subset being M_JWherein J ═ { J1, L, J_nH, where n is the minimum number of subsets; assuming that the scene regression coordinate of f (ω) is a PnP problem, four scene coordinates are sufficient to define a unique camera pose, so that n is 4, and since the result of randomly selecting four points for prediction at a time may be wrong, by randomly selecting 4 pairs of images to predict scene coordinates, a set of pose predictions f (ω) of n images may be finally generated, where each generated scene coordinate hypothesis f (ω) depends on a parameter ω;

s11-3, selecting an optimal pose hypothesis, selecting the optimal hypothesis by an evaluation mechanism for a hypothesis f (omega) with different scene coordinate regression performances, evaluating and determining the effect of selecting the camera attitude hypothesis and finally refining the hypothesis to generate a final estimation, wherein the method specifically comprises the following steps;

first, the reprojection error of pixel i of the image and the hypothesis f (ω) is defined as:

e_i(f,ω)＝||Cf^-1k_i(ω)-p_i||

in the formula p_iFor each pixel i of the image, the image coordinate (x)_i,y_i) C is the camera projection matrix, k_iIs an interior point, if e_i< τ, where τ determines an error threshold for the interior point;

scene coordinate regression work relies on pairingsInner point k_iCounting to score hypotheses, and in order to achieve end-to-end training of the neural network, constructing a differentiable function by using a sigmoid function:

in the formula e_iRepresenting a reprojection error, and controlling the softness of the sigmoid by using a hyper-parameter beta;

the evaluation function p (f) scores the agreement of each hypothesis with all scene coordinate predictions, hypothesis f, index j_j(ω) is selected based on a probability distribution P (j, ω, α) derived from the score value, hypotheses with high scores are more likely to be selected, and the final hypothesis is selected based on the softmax distribution P:

the size of the interior point counting score can be greatly different according to the difficulty of scenes, usually has large number of grades different under different environments, the evaluation score is kept in a reasonable range, the evaluation score is important for having wide distribution P (j; omega, alpha) and is important for stable end-to-end training, manual setting of the hyper-parameter alpha of each scene is a tedious task, alpha is automatically adapted in end-to-end training, and the size of the hyper-parameter alpha is selected through entropy:

training in end-to-end training is by argmin_α|S(α)-S^*Determining a parameter alpha by gradient descent, wherein the selection is to establish a target entropy in the just-started end-to-end training iteration and keep the target entropy stable in the whole process;

s11-4 refining the optimal pose assumption, the refining function R is an iterative process that uses the current pose estimate to determine and optimize inlier pixelsThe estimation is carried out alternately, if a neural network is used in the iterative process, end-to-end training must be realized, in order to improve the generalization and the precision of the model obtained by training, the training process is divided into two steps, the first step adopts an MAML method to divide the training process into an inner loop and an outer loop, so that the model has a better initial parameter, the second step can miniaturize the refining process, and finally the neural network model can output the final pose R (y)_j(ω))；

S21, obtaining the pose of the current person by performing scene coordinate regression on the image obtained by shooting through the mobile phone, wherein the coordinate regression method of the differentiable RANSAC is very high in precision, the error of the pose precision can be within 5cm and 5 degrees, and the position and the pose obtained through the image are used as the current initial position.

S22, obtaining the IMU data built in the mobile phone at the moment, carrying out reasoning and updating according to the obtained initial position posture data, namely the posture data at the previous moment, obtaining the posture data at the current moment,

s23, in the process of pose updating and navigation through the IMU, the pose obtained through the regression of the scene coordinates of the image is used as a new initial coordinate, the iterative updating process of the IMU is initialized, and the accumulated error of the inertial sensor is eliminated through high-precision visual positioning.

2. The method for mobile phone based navigation of airport underground parking spaces according to claim 1,

in the first step of the training process in step S11-4, the training process is divided into an inner loop and an outer loop by using the MAML method as follows:

introducing MAML initialization model parameters for model initialization and enabling the model to have better initial performance, dividing the training process of a neural network into an inner loop and an outer loop, wherein the inner loop realizes the training of basic model functions, the outer loop improves the training of model generalization, and a training set is also divided into two parts for the training of two loops;

when a 3D model exists in a training set of the inner loop, a rendered scene truth value is used for initialization, otherwise, a heuristic initialization is carried out by an approximate truth value, and the optimization of the inner loop is as follows:

the specific parametric iterative process of the inner loop is represented as:

in the formula, ω' is the optimal parameter of the inner loop iteration, ω is the initialization parameter, μ is the learning rate of the inner training process,

the gradient of the inner loop is indicated.

For the outer loop, the reprojection error calculated by the ground truth attitude is used, and a heuristic algorithm is used in the step of the inner loop, so that the correct depth of the scene point is effectively recovered;

therefore, the optimization process of the inner loop is as follows:

regarding the transfer of the inner loop and the outer loop parameters, a specific parameter transfer method is to perform element update or element optimization before the next batch of tasks is sampled, find the optimal parameter omega through the training of the inner loop, then calculate the gradient of omega for each inner loop, and update the random initialization parameter omega through a gradient update method, so that the random initialization parameter theta finds an initial parameter more similar to a target task, and a plurality of gradient steps are not needed when the next batch of tasks are trained;

the whole process is expressed as follows:

where ω is an initialization parameter, η is a learning rate hyperparameter of the outer loop,

is the gradient for the parameter ω' for the outer loop;

for the second training step, the model is trained in an end-to-end mode, all processes are required to be differentiable, including attitude optimization, and for the process of refining the hypothesis, the process of refining the pose is as follows:

because the argmin function is not differentiable, the differentiation of the thinning function is obtained by adopting a Gauss-Newton method:

in the formula, J_eRepresenting a Jacobian matrix comprising partial derivatives

Wherein h is_o＝R^t＝∞And (y) finally realizing end-to-end training.

3. A handset-based navigation method for airport underground parking space positioning according to claim 1, wherein each of said hypotheses y (ω) depends on parameters of the corresponding scene coordinates.

4. The method as claimed in claim 1, wherein the hyper-parameter α controls the width of the distribution, and the selection of the appropriate hyper-parameter facilitates end-to-end learning, which is to reduce the difference between the hypothesis scores so that the gradient values do not differ too far.

Images