Scholarlink

Abstract <p>As autonomous driving advances towards real-world implementation, accurately predicting the behavior of traffic participants remains a critical challenge for ensuring both safety and efficiency&#xa0;in driving. This study emphasizes the use of convolutional neural networks (CNN), employing long short-term memory (LSTM) and multilayer perceptron (MLP) as comparative models to investigate deep learning applications in behavior prediction. Utilizing Kaggle’s “prediction_for_validation_data.csv” dataset, the research develops a predictive system following data cleaning, feature extraction, and oversampling procedures. Experimental results indicate that CNN effectively extracts visual features through its convolution-pooling architecture, achieving an accuracy of 94.12% along with a perfect ROC-AUC score. In low-dimensional scenarios, MLP matches CNN’s accuracy by leveraging multilayer nonlinear transformations; however, it lacks depth in feature representation. Conversely, LSTM demonstrates limited performance due to its reliance on minimal temporal features, attaining only 52.94% accuracy—underscoring its dependence on rich temporal data. This study underscores the suitability of CNN for spatial contexts, highlights MLP’s efficiency in low-dimensional settings, and points out LSTM’s requirement for datasets abundant in temporal information. Future research may focus on optimizing CNN architectures, exploring cross-model fusion techniques, and developing multisource datasets to enhance predictive capabilities.</p>

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

Research on the Application of CNN in Autonomous Driving Behavior Prediction

  • 期刊论文

  • 发表日期：

    2026年1月1日

查看全文
  • Bingjie Zhao,
  • Changjie Wu,
  • Liang Zhao,
  • Maoxuan Peng

摘要

Abstract

As autonomous driving advances towards real-world implementation, accurately predicting the behavior of traffic participants remains a critical challenge for ensuring both safety and efficiency in driving. This study emphasizes the use of convolutional neural networks (CNN), employing long short-term memory (LSTM) and multilayer perceptron (MLP) as comparative models to investigate deep learning applications in behavior prediction. Utilizing Kaggle’s “prediction_for_validation_data.csv” dataset, the research develops a predictive system following data cleaning, feature extraction, and oversampling procedures. Experimental results indicate that CNN effectively extracts visual features through its convolution-pooling architecture, achieving an accuracy of 94.12% along with a perfect ROC-AUC score. In low-dimensional scenarios, MLP matches CNN’s accuracy by leveraging multilayer nonlinear transformations; however, it lacks depth in feature representation. Conversely, LSTM demonstrates limited performance due to its reliance on minimal temporal features, attaining only 52.94% accuracy—underscoring its dependence on rich temporal data. This study underscores the suitability of CNN for spatial contexts, highlights MLP’s efficiency in low-dimensional settings, and points out LSTM’s requirement for datasets abundant in temporal information. Future research may focus on optimizing CNN architectures, exploring cross-model fusion techniques, and developing multisource datasets to enhance predictive capabilities.