Adaptive Detector and Tracker on Construction Sites Using Functional Integration and Online Learning

Abstract

Tracking construction equipment is a major task when monitoring work in progress and performance on construction sites. Real-time location data of heavy equipment can be used not only to prevent collision accidents but also to predict work types and idle time. Many researchers have investigated the two-dimensional (2D) tracking of construction equipment from images. However, this method still frequently fails to track construction equipment in the long term due to the high interclass/intraclass variations of construction equipment and sites. In order to overcome this problem, this paper adapts and customizes a tracking method composed of two main concepts for (1) functional integration of a detector and a tracker and (2) real-time online learning using an automatically developed training database on site. The functional integration is first used to solve retracking issues and provide information used for database development. On the other hand, the online learning focuses on the use of a detector, which utilizes a site-customized database that is developed and updated automatically in real time. Validation was conducted using video stream data collected from four different construction sites. The average precision, recall rates, and data sampling accuracy were 86.53, 86.21, and 79.35%, respectively. The eigenvalues were also calculated as 0.66 and 0.39. The experiment results show the proposed method is able to consider the diverse characteristics of construction equipment and sites with promising performance. The contribution of this study is to improve performance and applicability of the functional integration and online learning for enhancing site awareness in the construction domain.