Automatic Object Detection-and-Counting Method Using Single-Shot Self-Supervised Learning and 3D Sensing in Resource-Limited Environments

Custom object detection systems are essential for meeting the specialized requirements of diverse applications, including manufacturing and biomedical engineering. Conventional deep-learning methods outperform these tasks but depend on large volumes of labeled data. This study presents a deep-learning method for object detection that reduces the amount of labeled training data to a single unlabeled 3D image obtained by a structured light system. By incorporating a 3D data augmentation stage and using a multi-scale prediction approach, we ensured robust detection and counting across different orientations and scales. This method drastically reduces the training data and improves the accuracy and robustness of the data processing using 3D sensors under limited environmental and computational conditions. The performance of this method indicates its suitability for adoption in a wide range of real-world sensor applications, providing a solid framework for advanced 3D sensor systems and offering a scalable and efficient solution for complex detection and counting tasks.