| Popis: |
The identification of vehicles from the images recorded by security cameras is now one of the most important investigative techniques. In the past, there were not variations in the shape of the side marker lights and the brake light so much, and it wasn't easy to distinguish the vehicle type from a nighttime rear image. Recently, however, LED-based devices have become popular, and the devices with unique shapes are increasing. For this reason, we focus on the nighttime rear vehicle images, and clarify how much the identification can be done. Nighttime rear vehicle images of 100 models, which is equivalent to 5% of the about 2,000 envisaged models, were used for the identification. The identification was performed with fine-tuning using ResNet-34 and ResNeXt-50-32x4d on PyTorch. The accuracy was 0.88 for the former and 0.90 for the latter. These results would be a first step to squeeze the vehicle type in the criminal investigation. |
