CN-120894762-B - Boundary frame prediction uncertainty modeling method based on residual error learning

Abstract

The invention discloses a boundary frame prediction uncertainty modeling method based on residual error learning, which comprises the following steps of firstly, constructing a boundary frame prediction uncertainty network based on a full connection layer and a residual error module; preparing track prediction data, calculating prediction confidence coefficient labels, training a prediction uncertainty network, and predicting network reasoning. The invention realizes the accurate quantification, efficient calculation and stable evaluation of the prediction uncertainty of the boundary frame through the network structure fused with the full connecting block and the residual block and the design of the uncertainty label, can remarkably improve the modeling precision and reduce the calculation cost, and can provide reliable prediction reliability evaluation for high-safety requirement application in the scenes of automatic driving, robot navigation and the like, thereby avoiding decision risk caused by the prediction uncertainty. The network has good generalization adaptation capability, and the modularized design enables the network to be seamlessly embedded into various existing track prediction networks, so that a flexible and efficient solution is provided for algorithm upgrading.

Inventors

• CHENG SHAOWU
• CHEN SIQIN

Assignees

• 哈尔滨工业大学

Dates

Publication Date

20260508

Application Date

20250716

Claims (6)

1. 1. The boundary frame prediction uncertainty modeling method based on residual error learning is characterized by comprising the following steps of: step one, constructing a boundary frame prediction uncertainty network based on a full connection layer and a residual error module; Step two, preparing track prediction data: Step two, acquiring historical track sequence data of targets in continuous frames, and recording position change information of each target in time sequence in detail; Secondly, generating a prediction boundary frame aiming at a subsequent frame by relying on the existing track prediction network, wherein the prediction boundary frame covers the position estimated information of a target in the next frame and is used as another important input of the network, and a complete input data pair is constructed together with the historical track sequence data; Step three, calculating a prediction confidence label: Step three, predicting the boundary box result Conversion to Wherein Is the upper left corner boundary coordinate, For the lower right corner boundary coordinates, calculating the root mean square error of the vertex coordinates of the prediction boundary frame and the true label ; Normalizing the root mean square error to a [0,1] interval; thirdly, converting the error value into a prediction confidence index through subtracting 1: step four, predicting uncertainty network training: By minimizing prediction uncertainty confidence Confidence with true uncertainty label The difference between the two realizes the training of the boundary frame prediction uncertainty network; Step five, predicting network reasoning: Prediction bounding box of input target Historical trajectories Confidence is output through a trained network 。
2. 2. The method for modeling the prediction uncertainty of the boundary box based on residual learning according to claim 1, wherein the boundary box prediction uncertainty network comprises a full-connection layer and a residual block, and the prediction boundary box result is firstly mapped to a feature space with a specific dimension through the full-connection layer: Wherein, the For the feature vector mapped to the bounding box, Is a weight matrix of the full connection layer, Is the weight of the bias and, Is the result of the prediction bounding box, Is a historical track; the output of the full-connection layer is used as input to continuously perform feature extraction and nonlinear transformation through two residual error modules, then one full-connection layer is connected, the features extracted by the residual error modules are integrated, mapped to an output space, and the confidence value of prediction is directly output 。
3. 3. The method for modeling uncertainty of boundary box prediction based on residual learning as claimed in claim 2, wherein in said residual module, input data is first passed through a full connection layer to obtain output Then, the ReLU activation function is used for processing the network, then a full-connection layer is connected, then a residual connection mechanism is used for adding the original input and the result after two-layer full-connection processing, and after the residual connection is completed, a nonlinear factor is further introduced through the ReLU activation function, so that the expression capacity of the network is enhanced.
4. 4. The residual learning-based boundary box prediction uncertainty modeling method according to claim 1, wherein the root mean square error of the prediction boundary box vertex coordinates and the true annotation The calculation formula of (2) is as follows: in the formula, And labeling the vertex coordinates of the bounding box for reality.
5. 5. The residual learning-based bounding box prediction uncertainty modeling method of claim 1, wherein the root mean square error normalization formula is: wherein the minimum value of the error is Maximum value of 。
6. 6. The residual learning-based bounding box prediction uncertainty modeling method according to claim 1, characterized in that the prediction uncertainty confidence Confidence with true uncertainty label The formula of the difference is as follows: Wherein: Is the loss function of the device, Is the number of samples in the batch; Is the network pair Uncertainty prediction of individual samples; Is the first True uncertainty confidence labels for individual samples.

Description

Boundary frame prediction uncertainty modeling method based on residual error learning Technical Field The invention belongs to the field of computer vision, relates to a boundary frame prediction uncertainty modeling method, and in particular relates to a boundary frame prediction uncertainty modeling method based on residual error learning. Background In the field of computer vision, trajectory prediction models are widely used in scenes such as automatic driving, robot navigation, and the like. In the prior art, most trajectory prediction models only output deterministic results, such as bounding box coordinates of the target location directly. However, such models lack a quantitative assessment mechanism for prediction reliability, and cannot inform the reliability degree of the prediction result. In practical application, if the prediction result output by the model has higher uncertainty but is not perceived, serious consequences such as misjudgment of road conditions, missing of important targets and the like may be caused by an automatic driving system. At present, although partial researches try to evaluate the uncertainty of a prediction result, the existing method has defects in network structure design and training strategies, and accurate modeling and quantification of the prediction uncertainty of the boundary frame are difficult to quickly and accurately realize. For example, the Monte Carlo dropout (Monte Carlo Dropout) -based method simulates uncertainty of a model by randomly discarding neurons in a network for a plurality of times, but relies on a large amount of forward propagation calculation, so that the efficiency is low, the Bayesian neural network-based method introduces uncertainty of prior distribution modeling, however, the accurate Bayesian inference calculation complexity is extremely high, the uncertainty estimation deviation is large due to common approximate inference, the uncertainty is acquired by training a plurality of independent models based on the ensemble learning method, a large amount of calculation resources and storage space are consumed, and the variability among models is difficult to control, so that the evaluation result is unstable. Disclosure of Invention Aiming at the problems that the existing track prediction model only outputs a deterministic result, a quantitative evaluation mechanism for prediction reliability is lacking, and the existing uncertainty evaluation method has the defects of low calculation efficiency, large reasoning deviation, high resource consumption, unstable evaluation result and the like on network structure design and training strategies, so that the boundary frame prediction uncertainty cannot be accurately modeled and quantized, the invention provides a boundary frame prediction uncertainty modeling method based on residual error learning. The method provides a boundary frame prediction uncertainty network fused by the full connecting block and the residual block, designs an uncertainty label calculation method, has outstanding advantages in improving calculation efficiency, accurately capturing uncertain characteristics and evaluating accuracy, can efficiently and accurately realize boundary frame prediction uncertainty modeling and quantification, helps to know prediction reliability degree, and avoids serious consequences caused by undetected prediction uncertainty. The invention aims at realizing the following technical scheme: A boundary frame prediction uncertainty modeling method based on residual error learning comprises the following steps: Step one, constructing a boundary box prediction uncertainty network based on a full connection layer and a residual error module: the boundary box prediction uncertainty network comprises a full connection layer and a residual block, and a prediction boundary box result is firstly mapped to a feature space with a specific dimension through the full connection layer: Wherein, the For the feature vector mapped to the bounding box,Is a weight matrix of the full connection layer,Is the weight of the bias and,Is the result of the prediction bounding box,Is a historical track; the output of the full-connection layer is used as input to continuously perform feature extraction and nonlinear transformation through two residual error modules, then one full-connection layer is connected, the features extracted by the residual error modules are integrated, mapped to an output space, and the confidence value of prediction is directly output ; In the residual error module, input data firstly passes through a full connection layer to obtain outputProcessing the network by using a ReLU activation function, then connecting a full connection layer, adding the original input and the results after two layers of full connection processing by using a residual connection mechanism, and further introducing a nonlinear factor through the ReLU activation function after the residual connection is completed, s