CNN-SVM-based Human-Computer Interaction Model for Automotive Systems in Complex Driving Environments

Abstract

In the complex driving environment, with the increase of task difficulty, the change of diversity and relevance, the phenomenon of perceptual mode conflict, strong cognition or increased difficulty of operation appears when drivers deal with tasks, which affects the execution of primary and secondary tasks. The information expressed and transmitted by multimedia technology is real-time, and only with real-time can we interact and transmit information with users. This paper discusses the design of automobile man-machine interaction based on multimedia information acquisition technology in complex driving environment. Based on users' situational awareness, this paper studies users' interactive needs and experiences in different driving situations, and proposes CNN-SVM (Convolutional Neural Network-support vector machine) emotional perception model. After automatically extracting spectral features through CNN, support vector machine (SVM) is used instead of traditional Softmax classifier to achieve accurate classification of multi class emotions. The experiment focuses on identifying the following core emotion categories: Anger, Neutral, Joy, and Anxiety, and verifies the model's generalization ability in driving scenarios through cross validation. In the spectrum segmentation experiment, the same network structure as CNN Net was used, and SVM was used instead of softmax classifier. CNN Net was used for training each time. After training, use the test sample set to calculate the input features of the softmax classifier, and input the new features into SVM to calculate the classification result of CNN-SVM. Select 500 test set images from the CityScapes dataset for testing, with MIoU (Mean Intersection Point on Consortium) used as the testing metric. The experimental results show that the model has improved the segmentation accuracy of roads, sidewalks, buildings, etc. Compared with several mainstream segmentation algorithms currently available, this algorithm has relatively small improvements for smaller target objects such as lights, signs, vegetables, and riders, with improvements of 1%, 0.5%, and 0.9%, respectively. In driving scenarios, users' judgment of "safety" mainly depends on whether secondary tasks occupy the cognitive and interactive channels of the primary driving task. This article explores how avoiding these two factors can provide users with a sense of security and improve their interaction experience.