Optimizing Driver Assistance Systems for Real-Time performance on Resource Constrained GPUs

Abstract

The importance of Advanced Driver Assistance Systems has increased tremendously due to their ability to reduce road fatalities by facilitating drivers for appropriate action selection in circumstances involving high probability of collisions. One of the major factors contributing to accidents on road is driver distraction and drowsiness. A variety of algorithms including several Forward Collision Warning algorithms have been proposed to alleviate the issue to road accidents. These algorithms are promising approaches to mitigate this problem. However, most of these proposals are computationally complex algorithms and require powerful GPUs to perform in real-time. Such GPUs are not only expensive but also have high power consumption. Thus, it is necessary to yield real time performance on resource constrained GPUs like NVIDIA's Jetson TX2 which is not only one of the most eminent GPU-enabled platforms for autonomous systems but also cost effective and power efficient [1]. This paper proposes utilization of pruning of Neural Networks and TensorFlow TensorRT to optimize computationally complex algorithms utilized for Driver Assistance Systems to obtain real-time functionality on TX2 without compromising the accuracy of the system.