基于RGB-D图像与头肩区域编码的实时人流量统计

Real-time Counting of Pedestrians Based on Head-shoulder Feature Coding and RGB-D Image

为解决基于彩色摄像机的人流量统计存在精度低、易受光照影响等问题,研究了利用RGB-D图像与头肩区域编码进行实时人流量检测的方法。在对RGB-D图像前景检测与空洞修复的基础上,采用随机点和漫水填充法提取潜在目标区域,并度量随机点与区域内极值点间的欧式距离以剔除噪声点的干扰,减少了冗余信息的计算量。根据人体头肩区域的先验信息,提出了一种七维头肩特征描述符对潜在目标区域编码,从而避免人体运动过程中产生的头肩部形变对头部区域检测的影响,采用SVM和阈值限制得到满足区域特征的头部区域。根据图像中不同区域的圆形度排序的思路,提出了头部区域筛选方法以去除无效区域,并构建了一种帧间目标匹配算法和多条计数线的方法,对已获得的头部区域实现计数,计数后获取人的身高数据。实验结果表明,该算法在教学楼楼道口和楼门口场景中人流量统计正确率高达99.30%以上,身高测量平均相对误差在2.59%以内,处理单帧图像的平均时间仅为18.31 ms。

Pedestrian counting methods based on color video camera have problems of low detection accuracy and difficult to identify people in uncontrolled illumination conditions.To solve these problems,this paper proposes a novel method of real-time pedestrian counting which utilize RGB-D image and head-shoulder feature coding.After foreground detection and holes repairing,potential object region is captured by random points and flood filling algorithm based on RGB-D image.Then,the Euclidean metrics is applied to measure spatial distance between the random point and the extreme point in the region,and thus eliminate the interference of noise points,which is utilized to reduce unnecessary computation.According to the prior information of head-shoulder region,a 7-dimensional head-shoulder feature descriptor is proposed to encode the potential target region,and SVM and threshold restriction are used to further obtain the head region with the regional characteristics,which avoid impacts of head-shoulder deformation during body motion.Based on the idea of sorting circularity of different regions in the images,a method using head region filtering to remove invalid regions is proposed.Meanwhile,a target matching algorithm between adjacent frames and multiple counting lines are constructed to calculate the quantity of acquired head regions and obtain the height data.The results show that the proposed method can accurately detect and count pedestrians in the entrances and inside stairways of a teaching building with recognition rate is 99.30%,and the mean relative error of height measurement is less than 2.59%,the average time of processing a single image only takes 18.31 ms.