3D target detection deep neural network model

Abstract

The invention discloses a 3D target detection deep neural network model. The 3D target detection deep neural network model comprises a data acquisition module which is mainly used for acquiring 3D target data required by network model training; the data classification module is mainly used for carrying out pre-classification on the obtained 3D target data; a data training module; the data storage module is used for preferentially storing the trained data and taking the data as subsequent samples; and a sample set storage module. According to the 3D target detection deep neural network model, the data classification module classifies the data to be trained in advance, so that the situation that the training time is influenced due to differences of multiple pieces of data is avoided, and the processing efficiency of the deep neural network model is improved; and secondly, the sample set storage module takes the trained data as samples to carry out preferential storage for subsequent training, so that the data demand is reduced, and normal training is ensured at the same time.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a deep neural network model for 3D object detection. Background Art

[0002] As one of the core foundations of autonomous driving perception systems, 3D object detection can be widely used in path planning, motion prediction, collision avoidance and other functions. In 3D object detection, a deep neural network model is needed to train data. After searching, it is found that the deep neural network model device in the prior art is typically such as the publication number CN109214426A, a method for object appearance detection and a deep neural network model. The method for object appearance detection has the following specific steps: building a deep neural network model; the deep neural network model has the ability to distinguish appearance through cyclic training and learning; according to the above large number of cyclic training and learning, the deep neural network model gradually converges to obtain the optimal weight of each feature value; performing convolution operation on the appearance picture of the object to be detected, and obtaining the classification result according to the set probability interval value. The deep neural network model includes a training module, an evaluation module and a prediction module. Its main feature is to solve many problems existing in the manual formulation of the judgment rules for the classification of object appearance, and overcome the current problems of low efficiency of manual operation or low accuracy of traditional automatic judgment; at the same time, due to its sustainable self-iteration and upgrading, its recognition efficiency will continue to improve in theory, the module is simple, the hardware cost is low, and the application range is wide.

[0003] In summary, the existing deep neural network models take a long time to train. To address the above problems, existing equipment needs to be improved. Summary of the invention

[0004] The purpose of the present invention is to provide a deep neural network model for 3D object detection to solve the problem of long training time of the existing deep neural network model mentioned in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a 3D object detection deep neural network model, comprising:

[0006] Data acquisition module: The data acquisition module is mainly used to obtain the 3D target data required for network model training;

[0007] Data classification module: The data classification module is mainly used to pre-classify the acquired 3D target data;

[0008] Data training module: The data training module is mainly used to train the classified 3D target data;

[0009] Data storage module: The data storage module is used to preferentially store the trained data and use it as subsequent samples;

[0010] Sample set storage module: The sample set storage module is used to store the data samples obtained by the data storage module, and the storage of the sample set storage module is backed up in time sequence.

[0011] Preferably, the data classification module is completed by SVM, i.e., support vector machine algorithm. Although it is difficult to classify and solve many original data, the support vector machine algorithm can perform a nonlinear transformation on such data through a suitable kernel function in order to achieve the purpose of linear separability.

[0012] Preferably, the support vector machine algorithm in the data classification module usually needs to construct the kernel function in the following format:

[0013] k(x i +y i )=ω(x i )·ω(x j ).

[0014] Preferably, the data classification module generally adjusts its penalty parameters and kernel function parameters in the optimization support vector machine algorithm to obtain a higher prediction classification accuracy. During the adjustment, common methods include trying values ​​within a certain range and optimizing parameters through genetic algorithms and particle swarm networks.

[0015] Preferably, the core principles of the data training module include forward propagation, back propagation, activation function, optimization algorithm, automatic feature extraction, high expressiveness, scalability and training challenge.

[0016] Preferably, the forward propagation data of the data training module refers to starting from the input layer, passing through the hidden layer, and finally reaching the output layer; each layer of neurons calculates according to the output of the previous layer and the weight and bias of this layer to obtain the output of this layer, and back propagation refers to calculating the error between the predicted output and the actual output of the network, and backpropagating this error back to the network layer by layer, while updating the weight and bias of each layer to reduce the prediction error.

[0017] Preferably, the activation function of the data training module is in the hidden layer, and the activation function is used to introduce nonlinearity so that the neural network can process complex patterns; commonly used activation functions include ReLU, Sigmoid and Tanh; the optimization algorithm refers to the use of the stochastic gradient descent algorithm, i.e., the SGD optimization algorithm, to update weights and biases during the back propagation process.

[0018] Preferably, the stochastic gradient descent algorithm usually uses a fixed learning rate for training during the training process, that is,

[0019] g t =f(θ t-1 )

[0020] θ t =-η*g t

[0021] where g t is the gradient, and η is the learning rate used to adjust the size of the gradient effect.

[0022] Preferably, the idea of ​​momentum of the stochastic gradient descent algorithm is added to the algorithm to calculate the size and direction of the final parameters. After the momentum is introduced, the parameter update method is:

[0023] g t =f(θ t-1 )

[0024] m t =μ*m t-1 +g t

[0025] θ t =-η*m t

[0026] Among them, m t is the accumulation of the current momentum, μ is the momentum factor, which is used to adjust the importance of the momentum of the previous step for parameter update. The use of momentum can make the algorithm jump out of the trap of local optimal solution.

[0027] Preferably, the automatic feature extraction in the data training module refers to the automatic learning and extraction of useful features by DNN from the raw data, reducing the reliance on manual feature engineering; the high expressiveness refers to the multi-layer structure; the scalability refers to the increase in data volume and computing power; the training challenge refers to the gradient vanishing or gradient exploding problem caused by the deep structure, which affects the training effect of the model and is solved by batch normalization.

[0028] Compared with the prior art, the beneficial effects of the present invention are: the 3D target detection deep neural network model,

[0029] The present invention can effectively solve the problem of long training time of existing deep neural network models through the coordinated use of a data classification module and a sample set storage module. The data classification module pre-classifies the data to be trained, thereby avoiding the impact of huge differences in size, type or other aspects on the training time of multiple data, thereby improving the processing efficiency of the deep neural network model. Secondly, the sample set storage module preferentially stores the trained data as samples for subsequent training, thereby reducing the demand for original data while ensuring the normal progress of training. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a schematic diagram of the overall structure of a 3D object detection deep neural network model of the present invention;

[0031] Figure 2 A schematic diagram of the mathematical principle of a DNN forward propagation algorithm for 3D object detection deep neural network model of the present invention;

[0032] Figure 3 This is a schematic diagram comparing a momentum acceleration algorithm and a stochastic gradient descent algorithm for a 3D target detection deep neural network model of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] See also Figure 1-3 , the present invention provides a technical solution:

[0035] Embodiment 1

[0036] A 3D object detection deep neural network model, including:

[0037] Data acquisition module: The data acquisition module is mainly used to obtain the 3D target data required for network model training;

[0038] Data classification module: The data classification module is mainly used to pre-classify the acquired 3D target data;

[0039] Data training module: The data training module is mainly used to train the classified 3D target data;

[0040] Data storage module: The data storage module is used to preferentially store the data after training and use it as subsequent samples;

[0041] Sample set storage module: The sample set storage module is used to store the data samples obtained by the data storage module, and the storage of the sample set storage module is stored and backed up in chronological order.

[0042] The data classification module is completed through the SVM, i.e. support vector machine algorithm. The support vector machine algorithm is difficult to classify and solve for many original data, but it can perform a nonlinear transformation on such data through a suitable kernel function in order to achieve the purpose of linear separability.

[0043] In a further example, the support vector machine algorithm in the data classification module usually requires the kernel function to be constructed in the following format:

[0044] k(x i +y i )=ω(x i )·ω(x j ).

[0045] In a further example, the data classification module generally adjusts the penalty parameters and kernel function parameters in the optimization support vector machine algorithm to obtain a higher prediction classification accuracy. When adjusting, common methods include trying values ​​within a certain range, and optimizing parameters through genetic algorithms and particle swarm networks.

[0046] In further examples, the core principles of the data training module include forward propagation, back propagation, activation functions, optimization algorithms, automatic feature extraction, high expressiveness, scalability, and training challenges.

[0047] In a further example, the forward propagation data of the data training module refers to starting from the input layer, passing through the hidden layer, and finally reaching the output layer; each layer of neurons calculates according to the output of the previous layer and the weights and biases of this layer to obtain the output of this layer, and back propagation refers to calculating the error between the predicted output and the actual output of the network, and backpropagating this error back to the network layer by layer, while updating the weights and biases of each layer to reduce the prediction error.

[0048] Specifically, assuming that the selected activation function is σ(z), and the output values ​​of the hidden layer and the output layer are a, then for the attached Figure 2 In the three-layer DNN, using the same idea as the perceptron, we can use the output of the previous layer to calculate the output of the next layer, which is the so-called DNN forward propagation algorithm. It is more complicated to use algebraic methods to represent the outputs one by one, but it is more concise to use the matrix rule. Assuming that there are m neurons in the l-1th layer and n neurons in the lth layer, the linear coefficients W of the lth layer form an n×m matrix W l , the bias b of the lth layer forms an n×1 vector b l , the output a of the l-1th layer forms an m×1 vector a l-1 , the unactivated linear output z of the lth layer forms an n×1 vector z l , the output a of the lth layer forms an n×1 vector a l , expressed in matrix form, the output of the lth layer is: a l =σ(z l )=σ(W l a l-1 +b l ).

[0049] The DNN forward propagation algorithm uses several weight coefficient matrices W and bias vectors b to perform a series of linear operations and activation operations with the input value vector x, starting from the input layer and calculating backward layer by layer until the operation reaches the output layer, and the output result is value. Input: total number of layers L, matrix W corresponding to all hidden layers and output layers, bias vector b, input value vector x Output: output of the output layer.

[0050] Since the gradient descent method has three variants: batch, mini-batch, and random, in order to simplify the description, here we take the most basic batch gradient descent method as an example to describe the backpropagation algorithm. The difference lies in the selection of training samples during iteration.

[0051] Input: total number of layers L, number of neurons in each hidden layer and output layer, activation function, loss function, iteration step a, maximum number of iterations max and stopping iteration threshold ε, input m training samples; Output: linear relationship coefficient matrix W and bias vector of each hidden layer and output layer.

[0052] In a further example, the activation function of the data training module is in the hidden layer, and the activation function is used to introduce nonlinearity so that the neural network can process complex patterns; commonly used activation functions include ReLU, Sigmoid and Tanh; the optimization algorithm refers to the use of the stochastic gradient descent algorithm, namely the SGD optimization algorithm, to update weights and biases during the back propagation process.

[0053] In a further example, during the training process, the stochastic gradient descent algorithm usually uses a fixed learning rate for training, that is,

[0054] g t =f(θ t-1 )

[0055] θ t =-η*g t

[0056] where g t is the gradient, and η is the learning rate used to adjust the size of the gradient effect.

[0057] In a further example, the idea of ​​momentum in the stochastic gradient descent algorithm is added to the algorithm to calculate the size and direction of the final parameters. After the introduction of momentum, the parameter update method is:

[0058] g t =f(θ t-1 )

[0059] m t =μ*m t-1 +g t

[0060] θt =-η*m t

[0061] Among them, m t is the accumulation of the current momentum, μ is the momentum factor, which is used to adjust the importance of the momentum of the previous step for parameter update. The use of momentum can make the algorithm jump out of the trap of local optimal solution.

[0062] In further examples, automatic feature extraction in the data training module refers to DNN automatically learning and extracting useful features from raw data, reducing dependence on manual feature engineering, high expressiveness refers to the multi-layer structure, and scalability refers to the increase in data volume and computing power; training challenges refer to the gradient vanishing or gradient exploding problems caused by deep structures, which affect the training effect of the model and are solved by batch normalization.

[0063] Embodiment 2

[0064] This embodiment is a further description of the above embodiment. It should be understood that this embodiment includes all the aforementioned technical features and is further described in detail.

[0065] The optimization algorithm of the data training module also includes the momentum acceleration algorithm, which can be regarded as an "adaptive" method. If the current gradient is consistent with the direction of the previous gradient, it means that we are on the right path and we will accelerate in this direction, just like when a ball rolls down from a high place, the greater the accumulated momentum, the faster the speed. If the current gradient is inconsistent with the direction of the historical gradient, it means that the current function changes dramatically, and we will slow down and adjust the direction. The main idea is to smooth the parameters in the network to reduce the amplitude of the gradient swing. The comparison diagram with the stochastic gradient descent algorithm is shown in the attached figure. Figure 3 As shown;

[0066] The optimization algorithm of the data training module also includes the Adam algorithm, which designs independent adaptive learning rates for different parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient; its process includes the following:

[0067] First:

[0068] V dω =0, S dω =0, V db =0, S db =0

[0069] Second: In the tth iteration, use the mini-batch gradient descent method to calculate dw and db

[0070] Third: Calculate the Momentum Index Weighted Average

[0071] Fourth: Update with RMSprop

[0072] Fifth: Calculate the corrected deviation of Momentum and RMSprop

[0073] Sixth: Update weights.

[0074] The Adam parameter configurations are: α learning rate / ; β1 exponential decay rate of the first-order moment estimate, such as 0.9; β2 exponential decay rate of the second-order moment estimate, such as 0.999. This hyperparameter should be set to a number close to 1 in sparse gradients (such as in NLP or computer vision tasks); ε is a very small parameter to prevent the denominator from being divided by 0 in the implementation to prevent it from being 0, such as 10E-8.

[0075] The Adam method can be seen as the culmination of momentum acceleration methods. On the one hand, it adjusts the forward movement method through momentum accumulation, and on the other hand, it uses the accumulated second-order information to adjust the step size.

[0076] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A 3D object detection deep neural network model, characterized in that: include: Data acquisition module: The data acquisition module is used to acquire the 3D target data required for network model training; Data classification module: The data classification module is used to pre-classify the acquired 3D target data; Data training module: The data training module is used to train the classified 3D target data; Data storage module: The data storage module is used to preferentially store the trained data and use it as subsequent samples; Sample set storage module: The sample set storage module is used to store the data samples obtained by the data storage module, and the storage of the sample set storage module is backed up in time sequence.

2. A 3D object detection deep neural network model according to claim 1, characterized in that: The data classification module is completed by the SVM, i.e., support vector machine algorithm. The support vector machine algorithm is difficult to classify and solve for many original data, but a suitable kernel function can be used to perform a nonlinear transformation on such data in order to achieve the purpose of linear separability.

3. The 3D object detection deep neural network model according to claim 1, characterized in that: The support vector machine algorithm in the data classification module usually needs to construct the kernel function into the following format: k(x i +y i )=ω(x i )·ω(x j )。 4. The 3D object detection deep neural network model according to claim 1, characterized in that: The data classification module generally adjusts the penalty parameters and kernel function parameters in the optimization support vector machine algorithm to obtain a higher prediction classification accuracy. During the adjustment, common methods include trying values ​​within a certain range and optimizing parameters through genetic algorithms and particle swarm networks.

5. The 3D object detection deep neural network model according to claim 1, characterized in that: The core principles of the data training module include forward propagation, back propagation, activation functions, optimization algorithms, automatic feature extraction, high expressiveness, scalability and training challenges.

6. The 3D object detection deep neural network model according to claim 1, characterized in that: The forward propagation data of the data training module refers to starting from the input layer, passing through the hidden layer, and finally reaching the output layer; each layer of neurons calculates according to the output of the previous layer and the weight and bias of this layer to obtain the output of this layer, and back propagation refers to calculating the error between the predicted output and the actual output of the network, and backpropagating this error back to the network layer by layer, while updating the weight and bias of each layer to reduce the prediction error.

7. The 3D object detection deep neural network model according to claim 1, characterized in that: The activation function of the data training module is in the hidden layer, and the activation function is used to introduce nonlinearity so that the neural network can process complex patterns. Commonly used activation functions include ReLU, Sigmoid and Tanh. The optimization algorithm refers to the use of the stochastic gradient descent algorithm, i.e., the SGD optimization algorithm, to update weights and biases during the back propagation process.

8. A 3D object detection deep neural network model according to claim 7, characterized in that: During the training process, the stochastic gradient descent algorithm usually uses a fixed learning rate for training, that is, g t =f(θ t-1 ) i t =-η*g t where g t is the gradient, and η is the learning rate used to adjust the size of the gradient effect.

9. A 3D object detection deep neural network model according to claim 8, characterized in that: The idea of ​​momentum in the stochastic gradient descent algorithm is added to the algorithm to calculate the size and direction of the final parameters. After the introduction of momentum, the parameter update method is: g t =f(θ t-1 ) m t =μ*m t-1 +g t i t =-η*m t Among them, m t is the accumulation of the current momentum, μ is the momentum factor, which is used to adjust the importance of the momentum of the previous step for parameter update. The use of momentum can make the algorithm jump out of the trap of local optimal solution.

10. The 3D object detection deep neural network model according to claim 1, characterized in that: The automatic feature extraction in the data training module refers to the automatic learning and extraction of useful features by DNN from raw data, reducing the reliance on manual feature engineering. High expressiveness refers to the multi-layer structure, and scalability refers to the increase in data volume and computing power. The training challenge refers to the gradient vanishing or gradient exploding problem caused by the deep structure, which affects the training effect of the model and is solved by batch normalization.

Images