Coarse-to-Fine Object Detection for Ride-Hailing Market Analysis

Abstract

To date, ride-hail services have developed a customer-centric platform to provide a positive experience for their customers. In this paper, we propose the computer vision techniques to extract market insight through integrated surveillance systems. To be specific, we classify the driver of ride-hail services that travel in a hundred routes according to their company in real-time. There are two major challenges to designing a real-time classification system: (1) almost similar in visual appearance between two classes of drivers, and (2) unbalanced sample distribution per class. In order to overcome these problems, in this paper, we introduce the use of the coarse-to-fine approach in the context of classifying drivers. We separate our approach into two main parts; weak object detection and refinement classification, respectively. As thoroughly evaluated in the experimental section, our approach can be used to analyze the CCTV data streams with high efficiency and robustness.