Real-Time Object Detection to Identify Adults and Children Using YOLO Algorithms

Remotely piloted vehicles such as drones or mobile robots are equipped with a light Graphics Processing Unit (GPU) to perform image processing tasks with limited complexity level. Thus, such systems essentially require light-weight algorithms to perform complex tasks such as object detection. For this purpose, the YOLO algorithms come in handy with different structures that support light, small, medium, and large devices like YOLO-tiny and YOLO-small. The aim of this research is to use a lightweight version of YOLO algorithms with high accuracy to detect and recognize three objects, namely child, female, and male, in real-time. To achieve this, we have used the YOLOv6-small and YOLOv7-tiny algorithms to detect the three objects. We have also included data augmentation techniques to enhance the detection accuracy and evaluation metrics (precision, recall, and MAP@5). The results have been promising with the YOLOv6-small, augmented model showing a +6% increase in MAP@5 for the child object and a +4% increase for the female object. On the other hand, we have managed to reduce wrong detection in some cases using the YOLOv7-tiny algorithm.