<![CDATA[The failure of the accident detection system on CCTV cameras can affect the increase in the death rate on the highway. The use of the CNN method in the construction of CCTV accident detection systems has been widely used before. However, common problems that are often encountered are dirty lenses and varifocal zooms that don't automatically focus, causing the quality of the resulting CCTV images to decrease, thus affecting system performance. In this research, a model was developed to detect accidents on CCTV images using the MobileNetV2 pre-trained model which was optimized by upscaling the dataset using the Real-ESRGAN model to produce more optimal performance. This study uses a CCTV image dataset totaling 989 and consisting of 2 types of prediction classes including accident and non-accident. The results showed that the MobileNetV2 model succeeded in producing 94% testing accuracy and an average inference time of 3.33 seconds in the GT test scenario. During the testing process, it was found that the model was not optimal if it identified new data with clustered objects. In addition, based on the test scenarios X2, X4, X8 it was found that the image quality calculated based on PSNR and SSIM values greatly influences classification performance such as accuracy, precision, recall, and AUC score.]]>
Copyrights © 2023
