A Human-Like Visual Perception System for Autonomous Vehicles Using a Neuron-Triggered Hybrid Unsupervised Deep Learning Method

Artikel i vetenskaplig tidskrift, 2024

Human-like visual perception systems are indispensable and vital components of human-like autonomous vehicles. In the real driving environment, there is much unlabeled information and the total number of categories of information is uncertain. While human brains are adept at processing such information, current methods are not. Thus, this study presented a novel hybrid unsupervised deep learning method to model the information processing mechanism of the driver’s visual perception. The proposed approach (CAE-SOM) was a neuron-triggered method, which leveraged the virtues of a convolutional autoencoder (CAE) and a self-organizing map (SOM) neural network. The CAE mimicked the hierarchical structures of the driver’s visual system to extract the high-level features, whilst the SOM neural network simulated the working principle of human brain neurons during the information judgment process to perform unsupervised clustering. The CAE-SOM method was built by using a dataset with eight common types of objects in road environments, and then it was tested on a public dataset LabelMe. The results showed that the CAE-SOM method performed well with an average accuracy of 90%. Compared with current unsupervised methods, the CAE-SOM model could improve the accuracy by nearly 10%. Compared with current supervised methods, this new model was still competitive, and its accuracy was close to the highest one. More importantly, the CAE-SOM model could reduce the cost of human labeling work in an unsupervised way and handle data from new categories that had never appeared. The outcomes could contribute to the visual algorithm optimization and safety improvement for autonomous vehicles.