Intent-Based Trajectory Prediction Approach Of Cyclists

Accidents between vehicle and cyclist have taken a large portion of traffic accidents in China.Cyclist motion predition is of great significance for both Advanced Driving Assistance System and Autonomous Driving,for protetecting the safety of cyclists and decreasing accidents rate.Current research in this field mainly focus on cyclist detection and tracking,while for cyclist motion prediction,the relevant research is still limited.Besides,most research failed to use cyclist pose and traffic environment information for prediction,and till now no integrated system is proposed to predict the cyclist motion.This thesis focuses on an integrated approach of cyclist intention inference and trajectory prediction method at unsignalized mix traffic intersections with interactions between cyclists and vehicles,incorporating the cyclist-related motion,pose and environment features.Firstly,the cyclist-related multi-features are extracted and classified against intersection scenarios,including motion layer,pose layer and traffic environment layer.The features as context cues are related to high-level behavior of cyclists at intersections(e.g.go straight versus turn left),and can be classified into dynamic environment cue,static environment cue and object cue.An algorithm for cyclist intention inference is designed based on Dynamic Bayesian Network which considers all the contect cues.The inference method is given for estimating posterior intent probability distribution.Secondly,based on the intention inference of cylists at intersections,an algorithm for cyclist trajectory prediction based on planning-based method and motion model is proposed.For the development of the prediction algorithm,a Frenet frame is introduced with key elements in traffic environment.The trajectories are generated with planning method by considering cyclist intentions and environmental constraints in 5-6s.Meanwhile,positions and velocities of cyclists within prediction horizon are predicted according to the motion model along the generated trajectories.Thirdly,the hardware and software platform are built for data collection of scenarios between cyclist and vehicle at intersections based on Robot Operating System(ROS).The image and Li DAR data are processed to build a cyclist database,for estimating the parameters in algorithm and verifying the performance of the algorithm.Finally,according to the characteristics of the interactions of cyclists and vehicles at intersections,12 scenarios are designed as the basis of verification of cyclist intention inference and cyclist trajectory prediction algorithm.The experimental results show the approach could in advance infer the cyclist intentions within 0.8s after the cyclist enters intersection.On the other hand,when considering relative distance between cyclist and ego vehicle,the intention is successfully estimated before the relative distance is less than 20 m.Based on the results of intent inference algorithm,the trajectory prediction method could generate reasonable cyclist trajectories in 5-6s and predict the future distribution of cyclist positions.The proposed approach connects the qualitative intent inference with the quantitative trajectory prediction domain to make a more accurate prediction within longer prediction horizon than conventional method,which is significant for intelligent vehicle on both VRU protection system and the module of path-planning.