Scholarlink

Autonomous driving at scale represents a safe and efficient platform for research in vehicle perception and control. This study compares the performance of four machine learning algorithms (Logistic Regression, Random Forest, Support Vector Machine (SVM) and Artificial Neural Network (ANN)) in classifying track type (straight or curved) for an autonomous vehicle on a scale of 1:10. Using a dataset of 2,672 samples, consisting of front camera images and steering angle values, eight statistical features were extracted to train the models. These were evaluated using metrics such as accuracy (accuracy), sensitivity (recall), F1 score and area under the curve (AUC), employing feature selection methods such as RFE, LASSO and Boruta. The results showed that Random Forest performed the best (accuracy: 0.96, AUC: 0.99), followed closely by SVM and ANN, while feature selection did not significantly affect the results. The findings laid the groundwork for future research in dynamic and more diverse environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparison of 1:10 Scale Training Models for Track Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bryan Betancourt-Ramírez,
  • Rafael Reveles-Martínez,
  • José G. Arceo-Olague,
  • Erika P. Sánchez-Femat,
  • Javier Saldívar-Pérez,
  • Carlos E. Galván-Tejada,
  • Luis C. Reveles-Gómez,
  • Manuel A. Soto-Murillo,
  • Antonio Martínez-Torteya,
  • José M. Celaya-Padilla

摘要

Autonomous driving at scale represents a safe and efficient platform for research in vehicle perception and control. This study compares the performance of four machine learning algorithms (Logistic Regression, Random Forest, Support Vector Machine (SVM) and Artificial Neural Network (ANN)) in classifying track type (straight or curved) for an autonomous vehicle on a scale of 1:10. Using a dataset of 2,672 samples, consisting of front camera images and steering angle values, eight statistical features were extracted to train the models. These were evaluated using metrics such as accuracy (accuracy), sensitivity (recall), F1 score and area under the curve (AUC), employing feature selection methods such as RFE, LASSO and Boruta. The results showed that Random Forest performed the best (accuracy: 0.96, AUC: 0.99), followed closely by SVM and ANN, while feature selection did not significantly affect the results. The findings laid the groundwork for future research in dynamic and more diverse environments.