一种改进的最大化AUC方法在障碍物检测中的应用

Improved Method of Maximizing AUC and Its Application to Obstacle Detection

在障碍物检测中,障碍物/非障碍物样本具有在特征空间上相互重叠范围大和分布不均衡的特点,而传统的分类器训练方法对这种数据的处理往往力不从心.针对此问题,文中提出一种改进的最大化ROC曲线下面积(AUC)方法来对分类器进行训练.该方法使用一种替代函数来作优化AUC的目标函数,同时将粒子群算法引入到AUC目标函数优化中,并通过使用巴特沃兹曲线和对适应值较差的粒子进行突变等方式对其进行改进.实验表明,使用该方法能够较好地解决因使用梯度法而产生的局部最优等问题,与已有的方法相比能更进一步提高障碍物的检测率,且算法本身可靠有效.

In the obstacle detection, obstacle/non-obstacle samples have characteristics of a large range of overlapping in the feature space and uneven distribution. The traditional training method for the classifier is not competent for dealing with such data. Thus, an improved method of maximizing area under the ROC （AUC） is proposed to train classifier. An alternative function is used as the objective function of optimizing AUC. Meanwhile, the particle swarm optimization is introduced to optimize the AUC objective function, and the particle swarm optimization algorithm is improved by using the Butterworth curves and particles with the low fitness value being mutated. The experimental results show that the proposed method effectively solves the local optimization caused by the gradient descent method. Moreover, the detection performance of the proposed method is improved compared with other existing algorithms, and the algorithm is reliable and efficient.