CenterNet: A Transfer Learning Approach for Human Presence Detection

Industrial robots have existed for decades, and their main design intention is to replace humans in those highly repetitive and dangerous tasks. Collaborative robots start to emerge in the industry as the market trends have moved towards the trend of high-mix low volume manufacturing. The concept of isolated working spaces is no longer ideal, and the acceptance issue of deploying collaborative robots by humans arises due to safety concerns. Strategic use of the sensors and various approaches have attempted to address the issue. However, in the light of artificial intelligence technology, limited studies are found to leverage the transfer learning technology in improving the safety of collaborative robots. Hence, the present study aimed to incorporate the use of visual sensors and investigate the use of transfer learning approach—fine-tuning to help detect the presence of humans at first. As the initial steps, a custom image dataset that comprised of 1463 images was acquired from the surveillance system in TT Vision Holdings Berhad. Image annotation was done; accordingly with bounding boxes and data splitting of training, validation and test sets with 70:20:10 ratio was performed. More variation of the images was introduced by applying data augmentation techniques such as hue, brightness and contrast adjustment. COCO pre-trained CenterNet object detection model was used as the source model to perform the transfer learning techniques. Relevant metrics and tools were used to monitor the model learning during the fine-tuning process and evaluate the robustness of the fine-tuned models. As a result, the CenterNet-ResNet101_V1_FPN fine-tuned model has achieved 81.9% with 73.28 FPS and the CenterNet-ResNet50_V1_FPN fine-tuned model with 80.57% AP with 42.30% FPS. These findings evidently describe the usage of human presence detection, suggesting that the transfer learning approach can be considered as the method that is used to improve HRI safety.