Transferability of a Convolutional Neural Network (CNN) to Measure Traffic Density

Whereas detecting individual vehicles in a video image using a convolutional neural network (CNN) prevails for traffic surveillance, CNNs also have been successfully adapted to counting vehicles via a regression method, which conveys the advantages of simplifying the model structure, and inference time can be reduced in the field. This model also demands much less human effort to tag images with labels. The number of vehicles in an image becomes the label, rather than bounding boxes drawn around every single vehicle. Nonetheless, the labeling task takes considerable time whenever a CNN model is trained and tested for a new road segment. There are two ways to alleviate the human effort involved in using this method. A previous study used a pseudo label pre-training method, and another study employed an image synthesis method to solve the problem. Besides these two methods, we investigated the model transferability to reduce the labeling effort. Using a CNN that was fully trained on images of a road segment, we devised a robust way to utilize the trained model for another site by transforming the model output with a simple quadratic equation. The utility of the proposed method was confirmed at the expense of a minute amount of deterioration in accuracy.

Automated summary

The study introduces a regression method using CNNs for vehicle counting, which simplifies model structure and reduces inference time; labeling images with vehicle count as labels requires less human effort, but training and testing on new road segments can be time-consuming; the study investigates pseudo label pre-training and image synthesis methods to alleviate human effort.