Towards Modeling of Perception Errors in Autonomous Vehicles

Detection and tracking of dynamic traffic objects such as pedestrians, cyclists, and surrounding ground vehicles is an important part of the perception of Autonomous Vehicle (AV). In practice, the presence of noise corrupts sensors' ideal performance, causing detection and state estimation of moving objects to be erroneous. These detection errors propagate through the overall system and potentially compromise the reliability and safety of the AV. To get an assurance that the vehicle will operate safely, any simulation platform for an AV must include a realistic representation of the fallacies in vehicle's perception. In this study, the perception error for a vision based detection algorithm of the camera sensor is modeled by applying auto-regressive moving average (ARMA) and nonlinear autoregressive (NAR) method. It will enable statistical error values to be injected into ideal values obtained from simulation models. The proposed approach is evaluated based on several test case scenarios using various environmental and traffic information. A comparative analysis of the behavior of the AV with and without perception error model for the imperfection of camera sensor has been undertaken using the CarMaker platform. The investigation of the impact on the behavior of the AV by the variation of the state (distance, brake-torque) clearly depict the effectiveness of incorporating the error model at detection level in CarMaker.