CN-122015816-A - Vehicle autonomous navigation method based on motion recognition and neural network assistance

CN122015816ACN 122015816 ACN122015816 ACN 122015816ACN-122015816-A

Abstract

The invention provides a vehicle autonomous navigation method based on motion recognition and neural network assistance, which comprises the steps of judging the motion state of a vehicle according to the output of an inertial measurement unit when satellite navigation is effective, and training a multi-head LSTM model under different motion states; when satellite navigation is invalid, the current motion state of the vehicle is firstly identified, the weights of all sub-models are set, then each sub-model of the multi-head LSTM model is predicted in parallel, and a fusion navigation result is generated by combining the weights. Therefore, the motion constraint and rules such as lateral zero speed are fully developed, the autonomous positioning accuracy of the vehicle without GNSS and other reference information is further improved, and the method has important practical application value.

Inventors

SUN YIHONG
HU HUAWEI
LI HAIQIANG
WANG HAIJUN
LI JIAHENG
ZOU SIYUAN
ZHUANG GUANGCHEN
GUO YUSHENG
HAN JIURUI
ZHOU YANAN

Assignees

北京自动化控制设备研究所

Dates

Publication Date: 20260512
Application Date: 20251218

Claims (10)

1. A vehicle autonomous navigation method based on motion recognition and neural network assistance, the method comprising: Designing LSTM submodels aiming at different motion states, wherein the LSTM submodels comprise a submodel 1, a submodel 2 and a submodel 3, the submodel 1 is a static prediction model and corresponds to a scene of vehicle parking and stationary, and model output is navigation system speed; acquiring inertial measurement unit data and executing inertial navigation calculation for inputting a parallel LSTM sub-model; And thirdly, acquiring satellite navigation data and executing combined navigation calculation when the GNSS is effective, and training an LSTM sub-model. Step four, identifying the vehicle motion state based on the output of the inertia measurement unit, wherein the vehicle motion state comprises parking static state, straight running and turning; Step five, based on the vehicle motion state identification result of the step four, using the inertial navigation result of the step two and the combined navigation result of the step three for training the multi-head LSTM sub-model of the step one; Step six, based on the multi-head LSTM sub-model trained in the step five, under the condition of GNSS invalidity, the inertial navigation data are used for predicting the vehicle speed information, and the method specifically comprises the following steps: The inertial navigation data are used for predicting each sub-model in the multi-head LSTM sub-model, and the prediction results of each sub-model are weighted and fused by combining the vehicle motion state recognition results of the fourth step to obtain higher-precision navigation information, wherein the static, straight-going and turning speed prediction results are respectively expressed as Sub-model weight values under different motion states and fusion results The following are provided: 1) Parking static: 2) And (3) straight running: 3) Turning: and step seven, based on the inertial measurement unit data acquired in the step two, taking the prediction result of the machine learning model in the step six as reference information, and performing integrated navigation by using Kalman filtering.
2. The autonomous navigation method of a vehicle based on motion recognition and neural network assistance according to claim 1, characterized in that for the sub model 1, the following design is performed: Model input, subtraction of harmful angular velocity components The output of the gyroscope, the quaternion corresponding to the carrier gesture and the navigation system acceleration deducting the gravity acceleration component Wherein, quaternion corresponding to the carrier gesture is calculated by using inertial device integration, and harmful angular velocity component Specifically, the device comprises an angular velocity component caused by the rotation angular velocity of the earth and an angular velocity component caused by the movement of a carrier on the curved earth surface; Model output, navigation system speed; the target standard is the navigation system speed calculated by inertial/satellite combined navigation; loss function, root mean square difference (RMSE) of model predicted navigation system speed and reference navigation system speed.
3. The autonomous navigation method of a vehicle based on motion recognition and neural network assistance according to claim 1, characterized in that for the sub model 2, the following design is performed: Model input, subtraction of harmful angular velocity components The output of the gyroscope, the quaternion corresponding to the carrier gesture and the navigation system acceleration deducting the gravity acceleration component Wherein, quaternion corresponding to the carrier gesture is calculated by using inertial device integration, and harmful angular velocity component Specifically, the device comprises an angular velocity component caused by the rotation angular velocity of the earth and an angular velocity component caused by the movement of a carrier on the curved earth surface; Model output, navigation system speed; the target standard is the navigation system speed calculated by inertial/satellite combined navigation; loss function, root mean square difference (RMSE) of model predicted navigation system speed and reference navigation system speed.
4. The autonomous navigation method of a vehicle based on motion recognition and neural network assistance according to claim 1, characterized in that for the sub model 3, the following design is performed: Model input, subtraction of harmful angular velocity components The output of the gyroscope, the quaternion corresponding to the carrier gesture and the navigation system acceleration deducting the gravity acceleration component Wherein, quaternion corresponding to the carrier gesture is calculated by using inertial device integration, and harmful angular velocity component Specifically, the device comprises an angular velocity component caused by the rotation angular velocity of the earth and an angular velocity component caused by the movement of a carrier on the curved earth surface; Model output, navigation system speed; the target standard is the navigation system speed calculated by inertial/satellite combined navigation; loss function, root mean square difference (RMSE) of model predicted navigation system speed and reference navigation system speed.
5. The method for autonomous navigation of a vehicle based on motion recognition and neural network assistance according to any one of claims 1 to 4, wherein the step two specifically comprises: collecting inertial measurement unit data, performing inertial navigation solution, subtracting the gyroscope output of harmful components The gesture quaternion Atti insQ calculated by pure inertial navigation. And calculates the navigational system acceleration f n , minus the gravitational acceleration component, based on Atti insQ and accelerometer outputs for input to the parallel LSTM sub-model.
6. The method for autonomous navigation of a vehicle based on motion recognition and neural network assistance according to claim 5, wherein the third step specifically comprises: And performing multiple GNSS effective sports car tests, performing inertial/satellite integrated navigation based on acquired inertial measurement unit data and GNSS data, and calculating a reference attitude quaternion Atti referQ ＝[q 0t q 1t q 2t q 3t ] T and a reference navigation system speed for training an LSTM sub-model.
7. The method for autonomous navigation of a vehicle based on motion recognition and neural network assistance according to any one of claims 1 to 6, wherein in the fourth step, a specific recognition strategy for recognizing a motion state of the vehicle based on an output of an inertial measurement unit is as follows: 1) Parking static: Calculating a 0.5 second average value of the angular velocity output by the three-axis gyroscope, continuously calculating a1 second standard deviation of the acceleration output by the three-axis accelerometer, continuously calculating a1 second standard deviation of the angular velocity output by the three-axis gyroscope, continuously calculating a2 second standard deviation of the angular velocity output by the three-axis gyroscope, and continuously calculating a 0.1 DEG/s of the angular velocity, wherein the three conditions are all met, and judging that the vehicle is stationary; 2) And (3) straight running: Calculating a heading angle of 5 seconds standard deviation, continuously calculating an average value of 0.5 seconds of angular speed output by the three-axis gyroscope, and calculating a heading angle of 5 seconds standard deviation, wherein the continuous 5 seconds standard deviation is smaller than 1.5 degrees, and calculating the average value of the angular speed output by the three-axis gyroscope to be smaller than 5 degrees/s; 3) Turning: The remaining cases of both cases 1) and 2) are not met.
8. The vehicle autonomous navigation method based on motion recognition and neural network assistance according to claim 7, wherein in the seventh step, a kalman filter state equation and a measurement equation are as follows: The state equation is X k ＝Φ k,k-1 X k-1 +w k-1 , wherein phi k,k-1 is a one-step transfer matrix from time t k to time t k-1 , w k-1 is the value of the system excitation noise sequence at time t k-1 , X k is the value of the state quantity at time t k , X k-1 is the value of the state quantity at time t k-1 , and X k comprises the northeast misalignment angle North east speed error δv n ＝[δV n δV u δV e ] T , latitude error Height error δh, longitude error δλ, carrier gyro drift epsilon b ＝[ε x ε y ε z ] T , carrier accelerometer zero bias v b ＝[▽ x ▽ y ▽ z ] T , carrier tie rod arm error L out , time delay T d ; The measurement equation is Z k ＝H k X k +V k , wherein Z k is a value measured at time t k , namely a value of a difference value between the gesture and speed information predicted by the six LSTM model and the speed and gesture information calculated by inertial navigation at time t k , H k is a value of a measurement matrix at time t k , and V k is a value of a measurement noise sequence at time t k .
9. The vehicle autonomous navigation method based on motion recognition and neural network assistance according to claim 1, further comprising the steps of executing the second to fifth steps when auxiliary information such as satellite navigation is available, updating the machine learning model online, and executing the second, third, fifth and seventh steps when auxiliary information such as satellite navigation is not available until the test is completed.
10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method of any of claims 1-9 when executing the computer program.

Description

Vehicle autonomous navigation method based on motion recognition and neural network assistance Technical Field The invention belongs to the technical field of inertial navigation, and particularly relates to a vehicle autonomous navigation method based on motion recognition and neural network assistance. Background The high-precision navigation positioning information is a precondition of vehicle path planning, safety obstacle avoidance and accurate control. At present, inertial/satellite combined navigation is mainly adopted for vehicle navigation positioning. However, the accuracy of satellite navigation is severely dependent on the validity of satellite signals, and in the case where a large number of satellite signals such as underground parking lots, tunnels, forest shadows and the like are invalid in the vehicle-mounted application scene, the satellite navigation signals are invalid, and errors of inertial navigation based on the product analysis accumulate over time. In particular, the microminiature low-cost inertial navigation system used for vehicle navigation has the defects of limited sensor precision and obvious error accumulation characteristic. Therefore, there is a need to design a vehicle-mounted autonomous navigation method that does not depend on external reference information such as satellite navigation. Disclosure of Invention The present invention aims to solve at least one of the technical problems existing in the prior art. Therefore, the invention provides a vehicle autonomous navigation method based on motion recognition and neural network assistance. The method can solve the problem that when the reference information such as satellite navigation is invalid, the accumulation of inertial navigation errors is difficult to meet the high-precision positioning requirement of the vehicle, can realize full-autonomous and high-precision navigation positioning of the vehicle independent of external reference information, and has important practical application value. The technical scheme of the invention is as follows: According to an aspect, there is provided a vehicle autonomous navigation method based on motion recognition and neural network assistance, the method comprising: Designing LSTM submodels aiming at different motion states, wherein the LSTM submodels comprise a submodel 1, a submodel 2 and a submodel 3, the submodel 1 is a static prediction model and corresponds to a scene of vehicle parking and stationary, and model output is navigation system speed; acquiring inertial measurement unit data and executing inertial navigation calculation for inputting a parallel LSTM sub-model; And thirdly, acquiring satellite navigation data and executing combined navigation calculation when the GNSS is effective, and training an LSTM sub-model. Step four, identifying the vehicle motion state based on the output of the inertia measurement unit, wherein the vehicle motion state comprises parking static state, straight running and turning; Step five, based on the vehicle motion state identification result of the step four, using the inertial navigation result of the step two and the combined navigation result of the step three for training the multi-head LSTM sub-model of the step one; Step six, based on the multi-head LSTM sub-model trained in the step five, under the condition of GNSS invalidity, the inertial navigation data are used for predicting the vehicle speed information, and the method specifically comprises the following steps: The inertial navigation data are used for predicting each sub-model in the multi-head LSTM sub-model, and the prediction results of each sub-model are weighted and fused by combining the vehicle motion state recognition results of the fourth step to obtain higher-precision navigation information, wherein the static, straight-going and turning speed prediction results are respectively expressed as Sub-model weight values under different motion states and fusion resultsThe following are provided: 1) Parking static: 2) And (3) straight running: 3) Turning: seventh, based on the inertial measurement unit data collected in the second step, taking the prediction result of the six-machine learning model as reference information, and performing integrated navigation by using Kalman filtering Further, for sub-model 1, the following design is performed: Model input, subtraction of harmful angular velocity components The output of the gyroscope, the quaternion corresponding to the carrier gesture and the navigation system acceleration deducting the gravity acceleration componentWherein, quaternion corresponding to the carrier gesture is calculated by using inertial device integration, and harmful angular velocity componentSpecifically, the device comprises an angular velocity component caused by the rotation angular velocity of the earth and an angular velocity component caused by the movement of a carrier on the curved earth surface; Model output, navigation system spe