AUTOMATED CREATION OF A SPECIALIZED DATA SET FOR CAR IMAGES

Abstract

The subject of this research is the algorithm and software for a system that automatically creates a specialized dataset of car images. The developed program runs on a Raspberry Pi 5 microcomputer, and the initial images are captured from a video camera. The software is written in Python and utilizes cloud technologies. The main tasks of the program are to capture initial images, select unique frames, and form a dataset from these frames. The frame selection process consists of three stages. The first stage involves global frame selection using perceptual hashing (pHash), which allows filtering out images that are generally similar to those already present in the dataset. In the second stage, frame selection is performed using car detection and image analysis, resulting in the extraction of frames containing new cars compared to those already in the dataset. Car detection in the images is performed using a YOLO convolutional neural network. The uniqueness of detected objects is determined using the Hungarian algorithm, taking into account the Intersection over Union (IoU) metric between bounding boxes of new and existing dataset images. The uniqueness of a car image is also determined by comparing the color values of car regions in the new image with those in the dataset. The third stage involves local selection of car regions using perceptual hashing, which results in frames containing only objects that are new to the dataset. During experimental testing, the developed system operated for 1.5 hours and captured 993 images, from which 121 unique frames (~12% of the initial set) were obtained after selection. The first stage provided the most significant data reduction – over 84%. Car detection was performed using a pre-trained YOLOv8m model (medium size), which ensures high detection accuracy. The resulting specialized image dataset was then used for fine-tuning faster YOLOv8n (nano size) model, which requires fewer computational resources. The final dataset is characterized by high diversity and suitability for further use in the fine-tuning of artificial neural networks.