CN-122023880-A - Parameter tuning method for automatic driving target classification model

CN122023880ACN 122023880 ACN122023880 ACN 122023880ACN-122023880-A

Abstract

The application provides an automatic driving target classification model parameter optimization method which is characterized by comprising the steps of preprocessing acquired multi-mode data of farmland environment, mapping super parameters of a target classification model into position vectors of crocodile individuals, constructing a multi-target fitness function, and iteratively optimizing the super parameters of the target classification model based on a crocodile algorithm and the preprocessed multi-mode data, wherein the crocodile algorithm comprises a position updating behavior mechanism, a submerging searching strategy and an escape strategy. The method and the device realize the improvement of the accuracy and the instantaneity of the target classification model in the embedded environment with limited resources.

Inventors

ZHOU JIE
CHEN WENBIN
Xian Yuting
SUN ZHENG
LI JIANGPENG
ZHANG YAO
Fei Hongmei
ZHANG CHUAN
MENG HUI
XIAO JING
HUANG NAIJUN

Assignees

石河子大学

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. An automatic driving target classification model parameter tuning method is characterized by comprising the following steps: preprocessing the acquired multi-mode data of the farmland environment; mapping the super parameters of the target classification model into position vectors of crocodile individuals, and constructing a multi-target fitness function; iteratively optimizing the super parameters of the target classification model based on an crocodile algorithm and the preprocessed multi-mode data; The crocodile algorithm comprises a location updating behavior mechanism, a submerging search strategy and an escape strategy.
2. The automatic driving target classification model parameter tuning method according to claim 1, wherein the multi-target fitness function satisfies: Wherein, the The value of the fitness is indicated as such, 、 And Respectively weighting factors and satisfying + + =1, 、 And Respectively representing classification accuracy, frame rate and computational complexity.
3. The automatic driving target classification model parameter tuning method according to claim 1, wherein the location update behavior mechanism satisfies the following calculation formula: Wherein, the Indicating the updated individual position of the crocodile, Indicating the current location of the individual crocodiles, Representing the location of the optimal crocodile individual in the current population, Representing a random step size of the step, , Representing a random disturbance that is indicative of the random disturbance, Representing the disturbance intensity.
4. The automatic driving target classification model parameter tuning method according to claim 1, wherein the update formula of the submergence search strategy satisfies: Wherein, the Indicating the updated individual position of the crocodile, Indicating the current location of the individual crocodiles, Representing the location of the optimal crocodile individual in the current population, Indicating the search depth, triggering the attack behavior when the value is 0.8 to 1.2.
5. The automated driving objective classification model parameter tuning method according to claim 1, wherein the escape strategy satisfies the following calculation formula: Wherein, the Representing the pseudo-locations of the optimal crocodile individuals in the current population, Representing the location of the optimal crocodile individual in the current population, Mean value of 0 standard deviation Is a gaussian noise of (c).
6. The method for optimizing parameters of an automatic driving target classification model according to claim 1, wherein the multi-mode data is acquired through a multi-spectrum camera and a radar sensor, and the preprocessing comprises illumination compensation, occlusion simulation and noise addition.
7. The method for optimizing parameters of an automatic driving target classification model according to claim 1, wherein the super parameters include a learning rate, a lot size, an anchor frame ratio, and a loss function weight.
8. The method for optimizing parameters of an automatic driving target classification model according to claim 1, wherein the target classification model adopts a modified MSPCNN-BiGRU architecture and comprises a main network and a self-adaptive anchor frame generator, wherein the main network adopts MobileNetV, the self-adaptive anchor frame generator is used for dynamically adjusting the anchor frame proportion based on the row spacing of crops, and the loss function is determined based on classification errors, bounding box regression errors and mask segmentation errors in a weighting mode.
9. The automatic driving target classification model parameter tuning method according to claim 1, characterized in that the method further comprises: the multi-mode data of the farmland environment collected in real time is input into an optimized target classification model, classification results of crops, weeds and obstacles in the farmland environment are output, and the optimized target classification model is deployed to an embedded platform of the agricultural machinery.
10. An automatic driving system for a vehicle, comprising a vehicle body, characterized by comprising the following steps: the multi-mode environment sensing module is used for collecting multi-mode data of farmland environment; an embedded platform for executing the automatic driving target classification model parameter tuning method according to any one of claims 1 to 9, and outputting classification results of crops, weeds and obstacles in a farmland environment based on the optimized target classification model; And the self-adaptive navigation and control module is used for dynamically adjusting the planned path of the agricultural machine based on the classification result, and the planned path is determined based on a path planning algorithm.

Description

Parameter tuning method for automatic driving target classification model Technical Field The application relates to the technical field of model super-parameter tuning, in particular to an automatic driving target classification model parameter tuning method. Background With the advancement of agricultural intellectualization, the application of automatic driving agricultural machinery in agricultural production links such as cultivation, seeding, fertilization and harvesting is gradually popularized, and the key point is to sense farmland environment in real time and make intelligent decisions. The target recognition model based on deep learning can accurately distinguish crops, weeds and obstacles, so that support is provided for automatic navigation. However, in embedded systems of agricultural machinery, the use of these models still faces a number of difficulties. Firstly, the agricultural machinery mostly adopts a low-power consumption embedded processor, the computing capacity of the low-power consumption embedded processor is limited, and the real-time reasoning of a complex deep learning model cannot be supported, so that the response speed of the system is influenced. Secondly, farmland environments have highly unstructured features, such as illumination changes, dynamic occlusion, earth reflection, etc., which may lead to insufficient robustness of the model, affecting classification accuracy. Finally, the super parameters (such as learning rate, batch size, convolution kernel size and the like) of the deep learning model are critical to performance, while the traditional super parameter searching method (such as grid searching and random searching) has large calculation cost and slow convergence speed, and is difficult to meet the requirements of agricultural automatic driving on instantaneity and calculation efficiency. Therefore, an efficient parameter tuning method for the automatic driving target classification model is urgently needed, and the accuracy and stability of the target classification model can be improved in an embedded environment with limited resources, so that the intelligent progress of agricultural machinery is promoted. Disclosure of Invention The application provides an automatic driving target classification model parameter optimization method, which is used for solving the defect that a target classification model deployed in an agricultural machine embedded system in the prior art is difficult to meet the requirements of agricultural automatic driving on instantaneity and calculation efficiency, and improving the accuracy and instantaneity of the target classification model in an embedded environment with limited resources. In a first aspect, the present application provides a method for optimizing parameters of an automatic driving target classification model, including: preprocessing the acquired multi-mode data of the farmland environment; mapping the super parameters of the target classification model into position vectors of crocodile individuals, and constructing a multi-target fitness function; iteratively optimizing super parameters of the target classification model based on crocodile algorithm and preprocessed multi-mode data; the crocodile algorithm comprises a location updating behavior mechanism, a submerging searching strategy and an escape strategy. In a second aspect, the present application provides an autopilot system comprising: the multi-mode environment sensing module is used for collecting multi-mode data of farmland environment; the embedded platform is used for executing the parameter tuning method of the automatic driving target classification model according to the first aspect and outputting classification results of crops, weeds and obstacles in the farmland environment based on the optimized target classification model; the self-adaptive navigation and control module is used for dynamically adjusting the planned path of the agricultural machine based on the classification result, and the planned path is determined based on a path planning algorithm. In a third aspect, the present application further provides an automatic driving target classification model parameter tuning device, including: The preprocessing module is used for preprocessing the acquired multi-mode data of the farmland environment; the mapping module is used for mapping the super parameters of the target classification model into position vectors of crocodile individuals and constructing a multi-target fitness function; The optimizing module is used for iteratively optimizing the super parameters of the target classification model based on the crocodile algorithm and the preprocessed multi-mode data; the crocodile algorithm comprises a location updating behavior mechanism, a submerging searching strategy and an escape strategy. In a fourth aspect, the present application further provides an electronic device, including a memory, a processor, and a computer program stored in