ADAPTIVE ENERGY-EFFICIENT MODELS FOR RECOGNITION OF ROAD INFRASTRUCTURE OBJECTS

The article presents the problem of recognizing the main road infrastructure objects (road sign, road marking, traffic lights). To solve it, it is proposed to use adaptive energy-efficient models that allow recognizing road objects taking into account dynamically changing conditions. A unique database of road scene images is proposed, which depict road infrastructure objects. The method of training and testing the adaptive energy-efficient model for recognizing road infrastructure objects includes the following successive stages: collecting initial data in the form of images of road scenes of recognizable objects (road signs, road marking, traffic lights); labeling the initial images; constructing the model architecture; selecting toolkits for modeling; training and testing the model; assessing the adequacy of the model using the error matrix and classification metrics. Data labeling was done manually using the LabelImg tool. Training, validation, and test samples were formed from the initial images of road scenes. The following tools were selected for training and testing the adaptive energy-efficient model: Python programming language, Google Colaboratory interactive cloud environment, Ultralytics library. Existing modifications of the models were trained: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x. The training time for the models was approximately 1.5 hours. The adequacy of the adaptive energy-efficient models is assessed using the error matrix and classification metrics (Precision, Recall, F1-score, mAP). In the future, it is advisable to expand the classes (temporary signs, barrier tapes, road barriers, traffic cones), use adaptive energy-efficient models in a mobile application or tablets to recognize the listed objects during repair work in real time.