Comparative Analysis of YOLOv8 and Mack-RCNN for People Counting on Fish-Eye Images

Monitoring the number of people in a building environment is key to achieving power efficiency, comfort, and safety in building use. The usage of fish-eye cameras in combination with neural network (NN) can provide a well-balanced compromise between price and measurement accuracy (ACC). Unfortunately, due to fish-eye camera distortion, solutions built for projective cameras may not work well. In addition, people can be counted with NN that solve different types of tasks, such as object detection or instance segmentation. As such, NN have different loss functions, and it is hard to forecast which approach could provide better performance for people counting on fish-eye images.To answer this question, the authors of this study compared YOLOv8 architecture vs Mask-RCNN under different training and pre-processing conditions. For this purpose, a custom data set was generated from different view angle cameras, and additional metrics were used apart from standard ones usually used for mask and bounding box evaluation. Experiments indicate that YOLOv8-N is a better solution as it requires fewer resources and at the same time provides top people count ACC of 95.