Algorithm for Three-Dimensional Detection of Pedestrians in Monocular Roadside View

Aiming to address the limitation of roadside monocular cameras in vehicle-road cooperative applications, which cannot directly capture the spatial morphology of pedestrians, we propose a pedestrian 3D detection algorithm based on the monocular roadside view, named Pedestrian3DKeypointNet. First, we construct a 2D-3D mapping matrix of the scene using an automatic traffic scene calibration algorithm. Then, we convert the known pedestrian 2D labeling information into 3D cylindrical labeling information to establish a 3D dataset for pedestrian detection. Second, we further extract features using an enhanced backbone feature extraction network and a global upper part module to obtain richer shallow and deep features. Finally, we introduce a new loss function to improve the convergence of training and enhance detection accuracy. Experimental results demonstrate that our improved pedestrian 3D detection algorithm increases detection accuracy by 7.66% compared to existing networks. It also achieves higher detection accuracy and speed compared to most current 3D target detection algorithms and effectively extracts the spatial morphology information of pedestrians in practical applications.