实时棒材图像识别与跟踪方法研究

多目标识别跟踪的关键问题是特征提取和目标匹配.为了提取生产线上堆积棒材的特征,提出粘连目标分割和多目标识别的方法.采用中值滤波和形态学滤波去除噪声,自适应阈值化和分水岭变换分割粘连目标;然后采用区域统计、参数识别、噪声区域去除以及聚类分析等手段进行目标特征识别,提取出棒材的质心点坐标作为特征;对棒材图像序列提出采用模板匹配、相近位移匹配和Kalman滤波的方法建立跟踪链,通过插入、删除、更新链节点进行目标跟踪;对于图像处理中可能出现的漏检目标和虚增目标,进行了计数结果校正.在现场采集了100帧连续图像后,采用此方法跟踪计数的精度为96.2%. 

Steel bar real-time recognition and tracking method

Feature extraction and pattern matching are the key problems in recognition and tracking system of multiple objects. In order to extract the features of stacked steel bars in real production environment, a method was proposed, which consisted of connected objects segmentation and multiple objects recognition. Median filter and morphological filter were applied in the steel bars image to remove the noise. Adaptive thresholding and watershed transform were used to segment the connected bar objects. Object centroid as the features was computed by means of regional statistics, parameter recognition, noise region removal and cluster analysis. For the image sequence of steel bars, the object tracking chain was established with template matching, near displacement matching and Kalman filtering. Target tracking was updated with inserting, deleting and refreshing of tracking chain nodes. The potential missing objects and false incremental ones were corrected in the counting result. At the production line 100 frames of sequential images were captured, and the tracking and counting method get the accuracy of 96.2%.