Driver’s Lane Change Intent Identification Based On Visual Characteristics And Vehicles’ Relative Movements

When lane change maneuver is executed, if the driver doesn’t observe the surroundingenvironment sufficiently, or wrong decision and judgements are made, this may easily lead toaccidents. The lane changing assistance system in the application serves turn signal as themain basis to identify drivers’ lane change intent, given the low usage rate and the short leadtime of the turn signal lamps, this result in poor performance of the assistance system. Ifdriver’s maneuver intention could be identified precisely before lane change occurs, and nolonger solely depending on turn signals as the main basis of identification, then we canevaluate the driving environments in advance, warn the drivers in the incubation period of thedanger to avoid traffic accidents, and then reduce the disturbance of traffic flow caused byunsafe driving behaviors.Combining with driving expectation and game theory, lane change decision-makingmechanism was thoroughly analyzed, related demand parameters of lane changing intentidentification were determined. Large scale real-world experiments were conducted on theselected routes, and16subjects took part in the experiments. Related parameters which couldreflect drivers’ visual search, operating characteristics, vehicles’ motion states and drivingenvironments were synchronously collected, so as to provide corresponding data items to lanechange intent identification research.Based on the deep analysis of driver’s fixation characteristics to the rearview mirrorbefore lane changing, time window which could characterizes driver’s lane changing intentwas determined, variation law of basic fixation and saccade parameters were analyzed withinthe time window. Regions of Interest were divided by the visual field plane, one-step andtwo-step transition probabilities between the regions were calculated with the application ofMarkov chain theory, the main visual search paths where the driver acquired informationduring the stage of lane keeping and lane changing intent were selected. Based on the Markovchain stationary distribution, glance proportions discrepancies among the interest regionsduring different driving stages were obtained. Combining with correlations analysis betweenhead rotations and gaze angles, as well as the analysis of driver’s head-eye coordination mode,visual parameters that can characterize driver’s lane change intention were extracted. In addition, based on the deep analysis of vehicle dynamics, drivers’ operation characteristicsand surrounding environment conditions, factors that own strong correlation with lanechanging intent and behavior were determined.Based on the evidence theory, identification framework was constructed, evidence chainthat could characterize drivers’ lane change intent was determined, based on generalizedhamming distance, basic belief assignment function was established. Sliding and accumulatedtime window was adopted, then multi-evidence fusion identification was carried out withinthe visual characteristics submodule. Given the selected factors ascertained, logisticregression models were built to identify lane change intent reflected by the vehicle dynamicsubmodule, whether a lane change would take place was further predicted. With theintegration of the two identification methods mentioned above, success rate and timesequence of the identification were soon afterwards determined. The following conclusionscan be drawn from this research work.1. Drivers’ low satisfaction in the current lane is the root of the lane changing intent.2. The time window that characterizes driver’s lane change intent is about five seconds.3. Fixation characteristics to the rearview mirror is the vital reflection of lane changeintent, in the lane changing intent stage drivers focus fewer minds on current lane than in thelane keeping stage.4. Under the condition of small amplitude transfer of fixation points, head rotations showweak correlation with gaze angles, while there is a strong correlation between them whenlarge scale transfer of fixation occurs, and head rotations can reflect drivers’ lane changingintent earlier than eye movements.5. Based on evidence theory and Logistic regression, drivers’ lane change intent caneffectively be identified, and accuracy rates are90.02%and86.07%, respectively.The research was sponsored by National Natural Science Foundation (51178053).