2nd ICAI 2022

International Conference on Automotive Industry 2022

Mladá Boleslav, Czech Republic

Figure 4: Architecture implementation scheme

Source: Own elaboration Described architecture implementation steps are complete and allows to train new classifier in less than 3 hours using CUDA® computing with GeForce® GTX 1650Ti. One of the key stages of implementation is the annotations in xml format and their subsequent conversion to YOLO format. Thanks to this, in case of insufficient results with YOLO, it is possible to use annotations in other architectures. After configuring the data generator, all layers of the model were frozen and at the same time the classifier trained on the UA-DETRAC dataset was removed. As configuration we use Keras RMSprop optimizer with learning rate = 2e-5, batch_size = 10 and epochs = 100. After training, the checkpoint with the least error was selected and implemented in OpenCV. Due to the high framerate in the input video and hardware limitations, decreasing of frame rate was also necessary. Completion of these steps allowed for the production implementation of the model with satisfactory efficiency. 3. Problem Solution The research was carried out using a commercial computer with a ninth generation Intel ® Core i7 processor and a GeForce ® GTX 1650Ti graphic card. As the development environment were used PyCharm, Keras library – in which the YOLO v3 network architecture was developed and the Django framework, in which the Web application was implemented. Application data is stored in PostgreSQL database, and annotations were made in the LabelImg application. As of the date of this article, all the mentioned applications are free of charge and allow for own implementation of the presented solution. The software design requirements outlined above directly influenced the conceptual scheme of the computer system designed to analyse transportation operations using object detection. Recording frames with the detection performed from the processed

113