Analysis of Differences in Street Visual Walkability Between Human and Machine Perception: A Case Study of an Anonymous University Campus

Recent studies of street visual perception have shown that street visual walk perception can be predicted using deep learning models, but machine-human perceptual differences limit the direct application of deep learning models to decision aids. In previous research, we developed a visual walk perception classification deep multitask learning (VWPCL) model for measuring visual walk perception (VWP). Also, the activation maps generated by the interpretable machine learning method Grad-CAM (Gradient-weighted Class Activation Mapping) were used to generate visual interpretations for the prediction results of the VWPCL model. However, its visualised machine perception results are not validated with real human visual perception data, and therefore designers have difficulty trusting the model’s perceptual prediction results. Based on this issue, this study conducted an experiment based on a desktop eye-tracker on a university campus to analyse the differences between human and deep learning models in street visual perception. The results of the study show that there are some differences between humans and deep learning models in street vision walkability perception. In future work, more quantitative analysis methods based on image comparisons and other sensory data will be incorporated for further in-depth research.