Development of real-time forward collision warning system using artificial neural network

Abstract

Most of the existing forward Collision Warning Systems (CWS) that involve the variable Perception-Reaction Time (PRT) have some negative effects on the collision warning performance due to the poor adaptive capability for the influence of different human PRTs. In an effort to address such issues associated with the parametric approaches, there has been a growing interest to develop a novel Artificial Neural Network (ANN)-based forward CWS in the field of Advanced Driver Assistance System (ADAS). However, since the previous studies show limitations for real-time learning to adjust and update pre-determined weights by considering spatiotemporal traffic pattern, they are of doubtful validity in the context of real-time forward CWS. This research develops a real-time learning rule called Rolling Period (RP)-based real-time learning algorithm that enables ANN-based forward CWSs to adjust their current weights depending on the real-time traffic information. With the real-time learning rule, Multi-layer Perceptron Neural Network-based forward CWS (MCWS) is proposed. The MCWS provides real-time forward collision warning by predicting potential critical deceleration in a subsequent few seconds. Comparison study demonstrates that the MCWS significantly reduces the number of nuisance and missed alarms compared to other conventional parametric forward CWSs, including Honda and Berkeley system, without any influence of the human PRT. To further enhance the performance of the MCWS, this study proposes Real-time associative memory-based forward CWS (RCWS), which updates and adjusts the pre-determined weights by using real-time sectional traffic information from vehicle sensor and smartphone in a cloud computing environment. Numerical analysis shows that the RCWS not only provides better warning performance compared to the MCWS, but also yields a significant improvement in the warning performance with a relatively low Market Penetration Rate (MPR) of in-vehicle sensors and communication devices. As an extension of the previous study on the RCWS, modified RCWSs using Multi-layer Perceptron Neural Network with two hidden layers and Radial Basis Function Network (RBFN) are proposed to mitigate impact of Communication Delay (CD) between smartphone and cloud server on the warning performance. Numerical study reveals that the overall performance improvements of the modified RCWSs compared to the original RCWS increases as the CD increases, which implies that the benefit of using the modified RCWSs increases as the time delay of uploading individual traffic data from the smartphone to the cloud server increases.