Comparative suspicious object detection in safe playgrounds using deep learning algorithms

According to the reports of Turkish Public Health Institution, approximately 250,000 rabies-risk animal bites occur per year in Turkey. Most of these bites are caused by dogs and most of the victims are the children who play in playgrounds in summer holidays. With the development of deep learning-based computer vision technology, autonomous detection of dangerous objects (handguns, knives, dogs, etc.) in these children's playgrounds has become a crucial security application. In this study, it is aimed to perform autonomous object (dog) detection by using fast emergency station located in the selected children's playground. Both traditional methods and region-proposal and single shot deep learning-based object detection algorithms have been examined. The evaluation scores of these algorithms are compared. After the comparison, in this thesis the YOLOv3 algorithm has been preferred due to its high speed and high mAP value. Single shot algorithms regress the image at once, estimate the class of objects and their position at the same time. Therefore, they are faster than region-proposal algorithms and more suitable for real-time detection. The YOLOv3 algorithm-based object detection model is proposed for children's playgrounds, due to the need for real-time and instant response in security systems and the fast moving of children and dogs in children's playgrounds. Firstly, in view of the problem of insufficient stray dog image data in the original datasets, new images of stray dogs have been taken from an animal shelter and they have been added to the dataset. Then, a new dataset that contains 878 images has been created by adding video recordings of the fast emergency station and images from an open images dataset. These new images have effectively enriched the diversity of training data and improved the training performance of the proposed model. The proposed model has been optimized by utilizing various hyperparameters and the results have been compared with each other. The model with the best evaluation scores is proposed as a new smart city security application and applied to detect dogs automatically by the fast emergency station located in the selected children's playground. The real-time application achieved 82.59% of AP @25 fps with adjusted hyperparameters to meet the speed and accuracy needs in the children's playgorund. Keywords: YOLOv3, Smart City, Children's Playground, Dog Detection, Fast Emergency Station