Autonomous Human Recognition System for Various Challenging Postures in Various Postures in Various Environments

Abstract

This paper introduces a deep learning method by using YOLOv3 as a medium to train human data samples and verify the presence of humans. The existing automated vehicles are designed and created for improving people’s everyday life. However, they have the limitation in term of safety technologies such as slow human detection speed and inaccuracy and unstable sensing capability [1]. Therefore, there is a need to improvise the current safety technologies by reducing the span time for recognition and more accurate data detection. This enhances the reliability of the system which will able to lower the accidence rate. This project aims to develop a reliable system, YOLOv3 that can recognize the human being with different postures indoors and outdoors environments. Besides, to evaluate the performance of YOLOv3 toward human detection in the image, video, and real-time. Lastly, to determine the best training last weight in performing human recognition. Google Colaboratory is used as a core training environment. It offers powerful Nvidia k80 and T4 with GPU memory 12 GB and 16 GB respectively that provided rapid training speed but there have two limitations which are maximum execution time is only 12 hours and maximum idle time is only 90 minutes. In order to detect the amount of light intensity surrounds the environment, the Arduino Uno R3 with LDR photosensor is used. As a result, the overall accuracy of the YOLOv3 human recognition model has achieved above 88 percent in bright environment. However, the system does not run efficiently when it is dark. In the future, more data could be used in the training to increase the accuracy of the YOLOv3 model. Further possible attempts also can include the use of an IR camera or a Learning-to-see-in-the-dark model with YOLOv3 to solve the problem of recognizing low-light images.