Improving Safety of Automated Driving by Minimizing Software Failures in Self-Driving Vehicles 2019-01-1052
Nowadays, Advanced Driver Assistance Systems (ADAS) play important roles in improving the safety of driving. ADASs along with new safety standards are now essential to improve the safety of automated driving. These standards differ from the safety standards in other industries because an autonomous vehicle rides in a very intricate stochastic environment where many details should be considered to maintain the safety. From one point of view, autonomous car failures are divided into two categories: hardware and software failures. According to the literature, if the reason of fault belongs to hardware, the troubleshooting will relatively be easier than software failure. In other words, if the fault is the subset of software, that would a more challenging task to detect potential fault and recover from the failure. In this paper, we propose a new approach to minimize the software failures, which may be unpredictable using a typical safety protocol. A predictive machine learning method is developed to improve decision making process in order to avert the potential accidents occurred by software defects. The proposed method enhances the safety and corresponding protocol by assessing objects detector's outputs using Deep Neural Networks. In other words, the system evaluates detected obstacle in the span of a specified number of frames to discern whether objects are detected correctly or not. Hence, it provides extra time to choose the suitable action before it reaches the danger, which includes lane changes, cut-in maneuvers, stopping the car, accelerating the car, slowing down, and allowing vehicles to pass. Results show that the new approach improve the safety by increasing the precision of failure detection.