Dwell Time Estimation Using Periodic Image Captures and Deep Learning

Abstract

The Innovative Truck Parking Availability System (iTPAS) employs computer vision algorithms (such as YOLOv8) to monitor truck parking occupancy in real-time at highway rest areas in Florida. Although iTPAS successfully detects instantaneous occupancy, it cannot determine how long vehicles remain parked, a key factor for real-time parking session analytics, including capacity planning and turnover estimation. In this work, we propose a dwell time estimation framework for iTAPS and similar systems using discrete periodic image captures, not continuous video. The main challenge is to classify whether vehicles appearing in consecutive images are the same or different. To this end, we collected periodic images from three counties in North Florida, extracted snippets with singular occupied zones, and created pairs of zones of either same or different vehicles. We then designed and experimented with two types of classification pipelines: zero-shot learning that needs no training at all, and a Siamese network that requires training. For both of them, we considered two backbone models: MobileNetV3 and Vision Transformer (ViT-B/16). Using our data, we observed zero-shot MobileNetV3-Large performed the best, with 93.80\% accuracy and 0.94 F1-score on the test set. With its much smaller size and less training time, zero-shot MobileNetV3-Large presents great potential in scalable deployment for dwell time estimation.