Enclosing contour tracking of highway construction equipment based on orientation-aware bounding box using UAV

Construction equipment tracking of highway construction site can obtain the spatiotemporal location in real time and provide data basis for construction risk control. The complete 2D moving of construction equipment in surveillance videos could be spatially represented by the translation, rotation and size change of corresponding images. To describe the temporal relationships of these variables, this study proposes a construction equipment enclosing contour tracking method based on orientation-aware bounding box (OABB), where UAV surveillance videos are employed to alleviate the occlusion problem. The method balances the rotation insensitivity of horizontal bounding box and the complexity of pixel-level segmented contour, which has three modules. The first module integrates OABB into a deep learning detector to provide detected contours. The second module updates OABBs with Kalman prediction to output tracked contours. The third module manages IDs of multiple tracked contours for construction equipment motions. Five in-situ UAV videos including 4325 frames were employed as the evaluation dataset. The tracking performance achieved 2.657 degrees in angle error, 97.523% in MOTA and 83.243% in MOTP.