Closed-loop cognitive information fusion method applied to unmanned driving

Abstract

The invention discloses a closed-loop cognitive information fusion method applied to unmanned driving, and the method comprises the steps: collecting original sensing information, carrying out the preprocessing, carrying out the weighted calculation, generating unified information fusion estimation, and carrying out the evaluation and scoring, thereby obtaining multi-angle uncertainty evaluation. A current tracking target and a multi-angle uncertainty evaluation result of the automatic driving vehicle are calculated, a possibility Q table of the next weight is generated according to the numerical value grid, and the total uncertainty corresponding to different selections of the Q table is calculated according to different weight possibilities and current measurement values pi of different sensors; and selecting the strategy in the Q table with the minimum total uncertainty. According to the method, the information fusion result is reversely transmitted to the beginning of the algorithm, the information collected by the sensor and the feedback information are fused at the same time, and self-correction of the information fusion algorithm is achieved, so that the estimation accuracy is remarkably improved, the accuracy of automatic driving planning control is improved, and the robustness and safety of the vehicle during autonomous operation are ensured.

Description

Technical Field

[0001] The present invention relates to a technology in the field of multi-sensor information fusion, specifically a closed-loop cognitive information fusion method applied to unmanned driving. Background Art

[0002] Information fusion is the core perception technology of unmanned driving. Current information fusion algorithms mostly use a one-way data flow method, which does not consider the specific problems of unmanned systems in dynamic operation, namely the uncertainty of the changing environment, which challenges the stability, robustness and safety of unmanned systems. At the same time, the one-way data flow method leads to the lack of feedback information, which fundamentally blocks the possibility of the algorithm self-correcting the information fusion results. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a closed-loop cognitive information fusion method applied to unmanned driving, which transmits the result of information fusion back to the beginning of the algorithm, and simultaneously fuses the information collected by the sensor and the feedback information to realize self-correction of the information fusion algorithm, thereby significantly improving the accuracy of estimation, enhancing the accuracy of automatic driving planning and control, and ensuring the robustness and safety of the vehicle during autonomous operation.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a closed-loop cognitive information fusion method applied to unmanned driving, comprising:

[0006] Step 1) After collecting and preprocessing the original sensor information, a unified information fusion estimate is generated through weighted calculation;

[0007] The original sensor information refers to: real-time point cloud information obtained by testing multiple radar sensors at different locations, as well as vehicle position information and speed information obtained by vehicle GPS, IMU and other sensors.

[0008] The preprocessing is to use the vehicle speed sensing algorithm to extract the position and speed of the vehicle in the world coordinate system from the point cloud information, and then combine the position information and speed information obtained by all different sensors into a vector p i =[x, y, z, v, θ], where: i represents the i-th sensor.

[0009] The vehicle speed sensing algorithm is implemented using but not limited to the technology described in "Vehicle Speed ​​Sensing Using Millimeter-Wave Interferometric Radar" by Klinefelter, Eric, and Jeffrey A. Nanzer (IEEE Transactions on Microwave Theory and Techniques, vol. 68, no. 12, 2020, pp. 4986-4996).

[0010] The unified information fusion estimation is obtained by weighted calculation of the vehicle speed and position calculated using multi-sensor sampling information, specifically: in, w i A real number between 0 and 1.

[0011] Step 2) Evaluate and score the unified information fusion estimate to obtain multi-angle uncertainty assessment: Total uncertainty in: represents the uncertainty of the position or velocity information measured or estimated by the i-th sensor, which is calculated by the weighted error sum of squares. i That is, the weight parameter in the above one-way information fusion, Σ i is the error covariance matrix of sensor i.

[0012] Step 3) By calculating the current tracking target of the autonomous driving vehicle and the multi-angle uncertainty evaluation results obtained in step 2, a Q table of the possibility of the next step weight is generated according to the numerical grid. For different weight possibilities and the current measurement values ​​p of different sensors, i Calculate the total uncertainty corresponding to different choices in the Q table, and select the strategy in the Q table with the smallest total uncertainty.

[0013] The present invention relates to a system for implementing the above method, comprising: an information processing and fusion module and a cognitive closed-loop analysis module, wherein: the information processing and fusion module collects raw data from different sensors and performs preprocessing and single-item information fusion to obtain a universal world coordinate system position and velocity vector; the cognitive closed-loop analysis module evaluates various uncertainties existing in one-way information fusion, uses a closed-loop adaptive adjustment module to construct weights in the next step of information fusion estimation, and acts on the information processing and fusion module at the next moment to achieve recursive reduction of uncertainty. Technical Effects

[0014] The present invention achieves structural reduction and control of estimation uncertainty and dynamic adaptive adjustment of the fusion algorithm by introducing a feedback loop. The result of information fusion is no longer solely dependent on the source of data and the accuracy of the fusion algorithm. Compared with the prior art, the present invention can handle various types of interference and uncertainty that may change dynamically in real time. On the one hand, the present invention utilizes a machine learning cognitive loop and no longer relies too much on the establishment of a physical model and its accuracy, as well as the adaptability of the fusion algorithm, greatly reducing the professional cost of modeling and fusion algorithm selection. On the other hand, the present invention can dynamically and effectively calculate and change the forward information fusion algorithm in real time, thereby greatly improving the estimation accuracy and algorithm robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the present invention;

[0016] Figure 2 It is a schematic diagram of the effect of the embodiment. DETAILED DESCRIPTION

[0017] like Figure 1 As shown, this embodiment relates to a heterogeneous T2T unidirectional information fusion method based on maximum likelihood method, including:

[0018] Step 1) After collecting different sensor information of the same observation object and preprocessing it, a negative log-likelihood function is constructed based on the preprocessed information and the estimated value obtained by the state equation, and the weights corresponding to different sensors are maximized to obtain the weighted matrix for fusing different sensor information in one-way information fusion, which specifically includes:

[0019] 1.1) Assuming that the measurement error follows a Gaussian distribution, the negative log-likelihood function can be constructed for the i-th sensor as Where: Σ i is the error covariance matrix of sensor i,

[0020] 1.2) Optimize weight w i Make the sum of all negative log-likelihood functions as minimum, specifically: So we get the weight matrix

[0021] Step 2) Constructing a feedback loop by fusing the one-way information obtained in step 1 through the Q-learning algorithm, specifically including:

[0022] 2.1) Make some changes to the information fusion algorithm, introduce a parameter matrix (initial value is the unit matrix) as feedback output in the unidirectional weighted matrix, and change the weight matrix in the fusion algorithm to the product of the feedback parameter and the original weighted matrix;

[0023] 2.2) In each iterative calculation step, the new forward information fusion result is input into the adapted Q-learning algorithm as the feedback controller, and the current information fusion result obtained in step 1 is scored by the standard Q-table obtained by the Q-learning algorithm to obtain the Q-value;

[0024] 2.3) The output parameter matrix obtained in step 2.2 is weighted to the weight matrix of the one-way information fusion and is transmitted back to the modified heterogeneous T2T algorithm based on the maximum likelihood method, realizing the real-time adaptive adjustment of the weight matrix of the information fusion algorithm.

[0025] After specific practical experiments, in the perception and control task of airport automatic berthing, when there are multiple sensors with different noise levels, the above method and the one-way information fusion control algorithm are run with the same noise parameters. The results are as follows Figure 2 As shown, the blue path is the unidirectional information fusion result without the introduction of Q learning, and the red path is the adaptive information fusion result of this method. It can be seen that the control result using this method has a faster adjustment response speed for noise and can converge to the target direction quickly.

[0026] Compared with the prior art, the present invention provides an adaptive self-correction capability for the information fusion mode of unidirectional data flow by introducing Q learning, which can effectively cope with the heterogeneous noise of different sensors and significantly improve the robustness and security of the algorithm.

[0027] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A closed-loop cognitive information fusion method for unmanned driving, characterized in that: include: Step 1) After collecting and preprocessing the original sensor information, a unified information fusion estimate is generated through weighted calculation; Step 2) Evaluate and score the unified information fusion estimate to obtain multi-angle uncertainty assessment: Total uncertainty in: represents the uncertainty of the position or velocity information measured or estimated by the i-th sensor, which is calculated by the weighted error sum of squares, where w i That is, the weight parameter in the above one-way information fusion, Σ i is the error covariance matrix of sensor i; Step 3) By calculating the current tracking target of the autonomous driving vehicle and the multi-angle uncertainty evaluation results obtained in step 2, a Q table of the possibility of the next step weight is generated according to the numerical grid. For different weight possibilities and the current measurement values ​​p of different sensors, i Calculate the total uncertainty corresponding to different choices in the Q table, and select the strategy in the Q table with the smallest total uncertainty.

2. The closed-loop cognitive information fusion method for unmanned driving according to claim 1 is characterized in that: The raw sensor information refers to: real-time point cloud information obtained by testing multiple radar sensors at different positions, as well as vehicle position information and speed information obtained by vehicle GPS, IMU and other sensors.

3. The closed-loop cognitive information fusion method for unmanned driving according to claim 1 is characterized in that: The preprocessing is to use the vehicle speed sensing algorithm to extract the position and speed of the vehicle in the world coordinate system from the point cloud information, and then combine the position information and speed information obtained by all different sensors into a vector p i =[x, y, z, v, θ], where: i represents the i-th sensor.

4. The closed-loop cognitive information fusion method for unmanned driving according to claim 1 is characterized in that: The unified information fusion estimation is obtained by weighted calculation of the vehicle speed and position calculated using multi-sensor sampling information, specifically: in, w i A real number between 0 and 1.

5. A closed-loop cognitive information fusion system for unmanned driving that implements the method described in any one of claims 1 to 4, characterized in that: include: Information processing and fusion module and cognitive closed-loop analysis module, where: the information processing and fusion module collects raw data from different sensors and performs preprocessing and single-item information fusion to obtain a universal world coordinate system position and velocity vector; the cognitive closed-loop analysis module evaluates the various uncertainties existing in one-way information fusion, and uses the closed-loop adaptive adjustment module to construct the weights in the next step of information fusion estimation, which acts on the information processing and fusion module at the next moment to achieve recursive reduction of uncertainty.

Images