Won Joon Yun¹, Soohyun Park¹, Joongheon Kim¹, David Mohaisen²
School of Electrical Engineering, Korea University, Seoul, Korea¹
Department of Computer Science, University of Central Florida, Orlando, FL, USA²
Guaranteeing real-time and accurate object detection, simultaneously, is of paramount importance in autonomous driving environments. However, the existing object detection neural network systems are characterized by a tradeoff between the computation time and accuracy, making it essential to optimize such a tradeoff. Fortunately, in many autonomous driving environment images come in a continuous form, providing an opportunity to use optical flow. In this paper, we improved the performance of YOLOv3-tiny, an object detection neural network that is designed to performs optical flow estimation in real time through FlowNet2-S. We improved the performance in terms of accuracy by 3.22%, and in terms of the number of detected objects by 59.6%. While the performance is improved, but because of the increased computation time, queue overflow may occur. To address this issue, we propose a Lyapunov optimization framework that observes the queue-backlog and selects the model every step accordingly.