驾驶人手机通话行为中基于图像特征决策融合的手势识别方法

为鲁棒检测自然环境中驾驶人的通话行为, 提出了一种驾驶人手机通话手势的识别方法。运用Adaboost算法检测驾驶人面部区域, 在YC_gC_r色彩空间中分别对面部肤色亮度分量和色度分量进行稀疏网格间隔采样, 由此建立了肤色的高斯分布模型; 针对驾驶室光照强度的不均匀性, 提出了肤色分量的漂移补偿算法, 建立了适应光照变化的在线肤色模型, 以准确分割左右手部肤色区域; 运用HOG算法获取手部肤色区域的2 376维HOG特征向量, 运用PCA方法将HOG特征降至400维; 同时提取手部肤色区域的PZMs特征, 并采用Relief算法筛选出权重最大的8个PZMs特征向量, 建立了融合PCA-HOG特征和Relief-PZMs特征的通话手势支持向量机分类决策。试验结果表明: 基于PCA-HOG特征的手势识别率为93.1%, 对光照变化的鲁棒性较好, 但易受到手部与头部转动的干扰; 基于Relief-PZMs特征的手势识别率为91.9%, 对于头部与手部姿态的耐受度较好, 但光照鲁棒性较差; 基于PCA-HOG和Relief-PZMs多元特征融合方法的手势识别率达到94.5%, 对光照波动、手部与头部转动等干扰条件具有较好的适应性。 

Hand gesture recognition method in driver's phone-call behavior based on decision fusion of image features

In order to detect drivers' phone-call behavior robustly in natural environment, a hand gesture recognition method was proposed. The Adaboost algorithm was used to detect driver's face region. In YC_gC_r color space, the brightness component and chroma component of facial skin were sampled by sparse grid, respectively, and a Gaussian distribution model of skin color was built. Considering the inhomogeneity of cab illumination, a skin color component drift compensation algorithm was proposed, and an online skin color model was established to adapt the changes of illumination, so that the skin color regions of right and left hands can be accurately segmented. The 2 376 dimensions HOG feature vector of hand skin region was extracted by HOG algorithm, and then PCA method was used to reduce HOG feature vector to 400 dimensions. Meanwhile, the PZMs features of hand skin region were extracted and 8 PZMs feature vectors with the largest weights were screened out by Relief algorithm. A support vector machine classifier decision for phone-call hand gesture was established based on the PCA-HOG and Relief-PZMs features. Experimental result shows that the hand gesture recognition rate based on the PCA-HOG features is 93.1%, and it has good robust to illumination changes but is easily disturbed by hand and head rotation. The hand gesture recognition rate based on the Relief-PZMs features is 91.9%, and it has good tolerance to head and hand gestures but has poor illumination robustness. The hand gesture recognition rate of the proposed multi-feature-fusion method combined with the PCA-HOG and Relief-PZMs is up to 94.5%, and it has good adaptability to illumination fluctuate, hand and head rotation, and other interference conditions.