| Title |
Dynamic Framerate SlowFast Network for Improving Autonomous Driving Performance |
| Authors |
(Byeong-Uk Jeon) ; (Kyungyong Chung) |
| DOI |
https://doi.org/10.5573/IEIESPC.2023.12.3.261 |
| Keywords |
YOLO; SlowFast network; Action recognition; Autonomous driving |
| Abstract |
Computer vision technology is used for autonomous driving and road traffic safety. Accordingly, studies on deep learning models that detect and analyze objects through images or videos are ongoing. On the other hand, deep learning algorithms that detect even the action of things require high computing performance. In addition, the computing performance of autonomous driving vehicles for processing such tasks is inferior. These classification processes are not used for recognizing and determining autonomous driving vehicles because it is impossible to process the classification of the action of autonomous driving vehicles in an autonomous driving vehicle on a real-time basis. This paper proposes a Dynamic Framerate SlowFast network for improving autonomous driving performance. Unlike pre-existing studies, the proposed model includes a cropping process through the YOLO model. In addition, it measures the similarity between unit frames through the SSIM and skips the input when the similarity exceeds a certain level. This process made it possible to reduce the number of frames entered into the model. Compared to the existing SlowFast Network, the performance evaluation compared the time required to analyze one image and the AUC of classification results when the number of input frames was reduced through similarity analysis techniques. The similarity analysis technique achieved the highest AUC when the SSIM was applied. The Dynamic Framerate SlowFast network proposed in this study achieved an AUC of 0.7126 and took an FPS of 0.7912 to analyze the entire verification video data. Compared to the pre-existing SlowFast network, which took an FPS of 0.5285 to achieve an AUC of 0.7531, the Dynamic Framerate SlowFast network achieved faster and more accurate results. Therefore, using the proposed technique, it is possible to achieve faster detection results while maintaining the object action detection AUC of the SlowFast network. |