论文部分内容阅读
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, the obstacle / non-obstacle samples have the characteristics of large overlap and unequal distribution in the feature space, and the traditional classifier training method often fails to handle such data.In response to this problem, An improved maximum area under the curve of AUC (AUC) method is used to train the classifier, which uses an alternative function to optimize the objective function of AUC and introduces the PSO into the optimization of AUC objective function It is improved by using Butterworth curve and mutation of particles with poorer fitness.Experiments show that this method can solve the problem of local optimization caused by the gradient method well, Compared to the method can further improve the detection rate of obstacles, and the algorithm itself is reliable and effective.