基于改进粒子群优化BP神经网络的弯道转向行为预测

Prediction of Steering Behaviors on Curves Based on BP Neural Network Optimized by Modified PSO

为研究驾驶人视觉特性和弯道转向行为的内在联系,借助模拟驾驶器,选取50名驾驶人在3种不同半径的弯道上进行驾驶试验.在整理采集的试验数据后,分别比较驾驶人视觉特性(瞳孔面积变化率、扫视速度、扫视幅度)、弯道转向行为(方向盘旋转率、车辆侧向加速度)与弯道半径之间的关系,并进一步提出一种以驾驶人视觉特性为预测因素,基于BP神经网络的驾驶人弯道转向行为预测方法.为使BP神经网络适用于小样本量的预测情况,需引入改进粒子群算法对BP神经网络进行优化.对粒子群算法的改进之处主要体现为:在粒子群算法进行搜索的过程中,采用动态惯性权值与自适应方法,解决了一般粒子群算法中粒子快速趋同的问题.在模型训练的过程中,选取BP神经网络的误差作为改进PSO算法的适应值,由事先确定的最大迭代次数与误差范围共同决定迭代的终止条件.最后,分别使用基于BP神经网络的驾驶人弯道转向行为预测方法,与基于改进粒子群优化BP神经网络的驾驶人弯道转向行为预测方法,对弯道转向行为进行预测,结果表明:基于改进粒子群优化BP神经网络的弯道转向行为预测方法相较传统预测方法具有更高的预测精度,可以有效地预测驾驶人弯道转向行为.

In order to study the inherent relationship between drivers visual characteristics and steering behaviors on curves, with automobile simulative instrument, 50 drivers are selected for driving simulation experiment on curves with 3 radii. After sorting out the collected experimental data, the relationships between driver’s visual characteristics ( change rate of pupil area, saccadic velocity, saccadic amplitude), turning behavior ( rotation rate of steering wheel, lateral acceleration) and turning radius are compared respectively. A prediction method of driver’s steering behavior on curves based on B P neural network with driver’s visual characteristics as predicting factors is proposed. In order to make BP neural network suitable for small sample size prediction, it is necessary to introduce MPSO algorithm to optimize BP neural network. The MPSO introduced the dynamic inertia weight and adaptive method to solve the problem of fast convergence of particles in general PSO during the process of searching. During the process of model training, the error of BP neural network is selected as the fitness value of MPSO , and the termination condition of the iteration is determined by the maximum number of iterations and the error range determined in advance. Finally, the prediction methods of drivers turning behaviors based on BP neural network and MPSO - BP neural network are used respectively to predict the steering behaviors of the driver. The result shows that the prediction accuracy of steering behavior prediction method based on MPSO - BP is higher than that based on traditional one, which can effectively predict the turning behaviors of driver on curves.