Deep Convolutional Neural Networks Based Vehicle Classification and Person Detection for CCTV Applications

The idea of forming life 'smarter' with next generation technologies in recent years, the focus on finding solutions to tasks in many sectors with Artificial Intelligence methods has increased, and new studies have been carried out to make these systems work robust with minimum error rate. These studies have shown that Convolutional Neural Networks achieves state-of-the-art performance in tasks such as object detection and image classification. In line with these updates, with this thesis project that we worked with Videotec company, we aimed to perform vehicle classification and person detection tasks in a low-cost way by using the Convolutional Neural Network model in CCTVs, which are available everywhere there is mobility. For this purpose, open source and private datasets were used to create a dataset pool for vehicle classification and person detection. Dataset balance was very critical, so both class balance and critical condition balance were considered while gathering images. During the model selection process, a choice was made among the state-of-the-art pretrained models in accordance with the scope of the project. Subsequently choosing the most appropriate one among trained models, the implementation process to CCTV is done. Although the FPS was low at real-time tests due to insufficient processor in the edge device, a promising step was taken with an accurate performance. In addition, on the cloud platform, the performance of the model was evaluated on the dataset that we created and our research represented that model performed successfully in critical conditions. A successful evaluation process was carried out to better measure the performance of the model, and a 92\% F1-score was obtained on our dataset. After the performance evaluation, the work was concluded with the real time detection test by connecting the Videotec cameras to the cloud platform via IP connection.