Pedestrian Detection Using Transfer Learning for Intelligent Vehicle Systems

Accidents are becoming more frequent these days as a result of the increase in the number of automobile vehicles on the roadways. Roadside pedestrians are frequently disregarded. Pedestrian deaths in traffic accidents are the most common. As a result, pedestrian safety is of the utmost importance. For providing safety features in intelligent vehicle systems (IVS), pedestrian detection is one of the most essential tasks. Adding a training module for pedestrian detection in existing IVS is however challenging. In this paper, various transfer learning techniques for the detection of pedestrians have been proposed. The different transfer learnings used in this work include MobileNet_v2, ResNet101, XceptionNet, DenseNet169, VGG16, and VGG19 along with their different classification metrics. These learning networks have undergone training and validation using a public dataset available on Kaggle and achieved comparable results. Among the various network, MobileNet_v2 and XceptionNet have outperformed the other methods and found suitable for pedestrian detection applications.