Anchor-free arbitrary-oriented construction vehicle detection with orientation-aware Gaussian heatmap

Effective detection of arbitrary-oriented construction vehicles is critical for ensuring construction site safety. Current construction vehicle detection methods are mostly anchor-based, which require complex manual setting for anchor proposals. This study proposes an anchor-free network for arbitrary-oriented construction vehicle detection with an orientation-aware Gaussian heatmap, which constructs a more appropriate intermediate state and provides more learnable information to accelerate training convergence and improve inference accuracy. The proposed network comprises feature extraction and regression parts. The former is used to extract multilevel features and restore the spatial information of construction vehicles, while the latter regresses the identification tuple including center point position, offset, width, height, and orientation angle in orientation-aware bounding boxes. Moreover, this study created a multiscale dataset captured from 18 different actual construction sites for training and verification, including 600 images and 1570 construction vehicles. Comparisons with different state-of-the-art methods (including anchor-based, anchor-free and segmentation-based methods) demonstrate the accuracy and effectiveness of the proposed method.