| Title |
Avoiding Performance Degradation of DNN based Safety-Critical Applications in Autonomous Vehicles |
| Authors |
김도영(Doyoung Kim) ; 김명선(Myungsun Kim) |
| DOI |
https://doi.org/10.5573/ieie.2020.57.9.56 |
| Keywords |
Deep neural network; Safety-critical; Multi-DNN; Priority scheduling; Latency |
| Abstract |
The rapid development of DNN technology is leading to outstanding performance improvement in autonomous vehicles. In this system, various DNN model-based applications such as object detection, lane departure prevention, drowsy driving prevention, and surrounding image analysis are performed. The applications hosting the DNN models are classified into safety-critical ones with high priority and others with medium or low priority not directly related to driving. Therefore, DNN accelerators such as GPUs are shared and used by DNN models with different priorities. In addition, since it is performed in the form of one batch process per DNN model, it is difficult to firstly allocate the DNN models used by safety-critical applications that arrive sporadically to the GPU. In order to solve this problem, this paper proposes a framework for requesting computation to the GPU for each layer of the DNN model based on priority. As a result of experiments after running the proposed technique on an actual off-the-shelf board, the performance of high-priority DNN models improved upto 43.9% compared to that of before. |