People Detection System Using YOLOv3 Algorithm

In crowd security systems, precise real-time detection of people in images or videos can be very challenging especially in complex and dense crowds whereby some individuals could possibly be partly or entirely occluded for varying lengths of time. Thus, this paper presents a large Convolutional Neural Network (CNN) that is trained using a single step model, You Only Look Once version 3 (YOLOv3) on Google Colaboratory to process the images within a database and to accurately locate people within the images. YOLOv3 splits the image up into regions and predicts bounding boxes and predicts the probabilities for each region. These bounding boxes are weighted by the projected probabilities and finally, the model is able to make its detection based on the final weights. This model will be using a customised dataset from Google's Open Images with 500 high resolution images. Once trained, the neural network able to successfully generate the test data and achieve a mean average precision (mAP) of 78.3% and a final average loss of 0.6 on top of confidently detecting the people within the images.