Detection of Construction Equipment Using Deep Convolutional Networks

Abstract

Vision-based monitoring methods have been investigated for understanding construction site contexts. However, detection capabilities of such methods are still insufficient to be utilized in general construction sites due to dynamic outdoor conditions and appearance variances of construction entities. To improve performance of a construction entity detector, we propose a detection method using a region-based fully convolutional network (R-FCN). R-FCN consists of two main parts, which are a fully convolutional network and a region proposal network. The fully convolutional network extracts hierarchical object features through a supervised learning process, while a region proposal network generates a set of object candidate regions in an image to localize target objects. To evaluate the generalization performance of the detection method, a benchmark dataset is collected from ImageNet for five classes (dump truck, excavator, loader, concrete mixer truck, and road roller), having various object appearances within a class in different backgrounds. A state-of-the-art performance, mean average precision of 95.61%, was recorded from the experiment. The proposed method shows a potential for the universal detector that can detect construction equipment on every construction site.