摘要
针对侧视角下司机体型差异引起的驾驶姿态识别误差问题,提出一种空间先验概率模型增强肤色直方图的司机姿态特征提取方法。首先,利用联合色彩空间进行训练数据的皮肤检测,并根据训练样本的统计结果建立先验概率模型,对皮肤检测结果进行加权计算;然后,以方向盘的中心为原点建立极坐标系,对局部扇形区域进行参数化数学建模,并利用遗传算法实现侧视角下司机姿态图像的空间分割最优化;最后,以最优空间分割结果作为姿态直方图特征的横坐标,从先验概率模型中提取司机姿态特征直方图。实验结果表明:经过先验概率模型加权的肤色直方图特征能够明显减少司机身材差异对姿态识别结果的影响,在相同分类器下,该方法比传统肤色直方图特征提取方法的平均识别率高10%以上;进一步地,将该方法与神经网络相结合,与其他常用姿态识别方法的平均识别率相比提高12%以上。
Aiming at the problem of driving posture recognition difference,a enhanced driver′s posture feature extraction based on the skin-color histogram of spatial prior probabilistic model is proposed.First,multiple color spaces are combined in driver skin-color segmentation,a prior model is proposed based on the statistical results of the test database,and a weighted calculation of the skin-color test results is conducted.Then,the polar coordinate system is established with the steering wheel center as the origin,and a parametric mathematical model is proposed for the local fan shape,then the genetic algorithm is employed to optimize the spatial segmentation of the driver′s posture image from the side view.Finally,the optimal space segmentation result is used as the horizontal coordinate of the posture histogram to extract the driver′s posture histogram with the help of prior probabilistic model.The results show that the skin-color histogram weighed by the prior probabilistic model can remarkably reduce the difference of drivers′stature on the posture recognition results.Under same classifier this method′s average recognition is 10%higher than the traditional skin-color feature extraction.Moreover,the average recognition rate under the combination of neural network and proposed method is 12%higher than other traditional methods.
作者
王晗
孙雨
尹杰
WANG Han;SUN Yu;YIN Jie(School of Transportation and Civil Engineering,Nantong University,Nantong 226019,China;School of Information Science and Technology,Nantong University,Nantong 226019,China)
出处
《南通大学学报（自然科学版）》
CAS
2020年第3期34-41,共8页
Journal of Nantong University(Natural Science Edition)
基金
国家自然科学基金项目(61872425)
江苏省高校自然科学基金面上项目(17KJB520029)。
关键词
驾驶姿态识别
空间先验概率模型
极坐标系
遗传算法
driver posture recognition
spatial prior probabilistic model
polar coordinates system
genetic algorithm
