Abstract
The motion status recognition of the preceding vehicle in a long-distance region is a requirement for autonomous vehicles to make appropriate decisions and increase their comprehension of the environment. At present, the lane change behavior of the leading vehicle at a short distance is detected using stereo cameras and LIDAR. However, the short detection distance (about 100 m) does not meet the requirements of high-speed driving of autonomous vehicles on expressways; this is a fundamental problem limiting the development of autonomous vehicles exhibiting human-like behavior. In this study, a comprehensive model consisting of a back-propagation (BP) neural network model optimized by a particle swarm optimization (PSO) algorithm and a continuous identification model is developed based on the results of naturalistic on-road experiments using millimeter-wave radar data. By considering different time-to-lane crossings (TLCs), the PSO-BP neural network model is trained using real vehicle lane change data and is implemented when the TLC of the leading vehicle is longer than 1.0 s. In contrast, when the TLC is less than 1.0 s, the continuous recognition model of the TLC is used. By comparison with BP neural network model, the recognition accuracy rate of the proposed model is increased from 80% to 87% after PSO optimization for a time window of 1.0 s; these results meet the recognition requirements of autonomous driving systems for distant targets. The findings of this study improve the cognitive competence and safety of autonomous driving systems.
Funding Information
  • National Key R&D Program of China (2018YFB1600501)
  • Changjiang Scholars and Innovative Research Team in University (IRT_17R95)
  • National Natural Science Foundation of China (51775053)

This publication has 63 references indexed in Scilit: