Project information
- Category: Machine Learning
- Data: Public transport passenger smart card data
- Project: Work of MSc student - Heqi
- Methods: Image Processing, CNN, RNN, Class Activation Maps, Transfer Learning
- Technologies: Python, pandas, keras, opencv
- Journal: Master Thesis
Project details
The study analyzes how deep neural networks identify spatiotemporal traffic patterns for traffic predictions and uses a pretrained Inception ResNet v2-based image classifier and Grad-CAM for the task. The Amsterdam freeway network is used as a case study. Despite expecting competitive performance, the model could not accurately capture dynamic traffic characteristics or provide reliable predictions. The results underline the influence of inductive bias, importance of fine-tuning, and model-data compatibility in deep learning models. These insights are beneficial for selecting appropriate models for future network-wide traffic speed prediction tasks.