Abstract:

The number of automatic systems that analyze images obtained from unmanned aircraft with artificial intelligence systems is increasing. We also see examples of these studies in urban and traffic practices. It is very important to determine the number of cars in the picture for these applications. The training of artificial intelligence systems requires as many labeled data as possible. It can be very difficult depending on the method in which the labelling process is done. In this study, a deep learning model is recommended that automatically determines the number of cars from IHA images. The sufficient point labelling data for the proposed model training makes the data set creation phase quite easy. A new loss function has been suggested for the training of deep models with point labelling. Experimental results on the CARPK data set have been made in order to assess the success of the proposed method. The results obtained compared to the models trained with frame labelling, it appears that the recommended method has achieved a level of success that can compete with these models.

Keywords: