Visual Metrology For The Pedestrian’s Height With Uncalibrated Video

Body behavior recognition is a very key task in criminal investigation, safety monitoring and behavior analysis. The height of a human is also an essential parameter for recognition. For example, in forensic cases, the suspects’height is a very important feature to describe the human body. When camera resolution used in video surveillance is not very high, or too far away from the target, the information of the human body we get, especially the facial features are often vague, so the only information we can use is the height. Compared with other measurement methods, such as ultrasonic, laser and radar, the measurement technology based on image/video just need to collect the image/video information, then use measuring software to measure offline or online at any time. So it has many advantages such as safety, repeatability and the cheap cost. At the same time, compared with measurement technology based on image, measurement technology based on video not only can use the motion information in video but also calculate the result using probability and statistics to improve the accuracy of measurement.The height measurement technology of video consists of calibration and no calibration measurement. Calibration measurement need a high precision special calibration blocks such as a checkerboard, to determine the camera parameters and distortion coefficients. The process is cumbersome, but the advantage is high precision. While no calibration measurement doesn’t need the calibration block, but utilizes fixed objects such as buildings, tables and chairs, or some structured information of moving objects such as parallel, perpendicular, collinear points in 3d scene to achieve the purpose of measurement. And the method is more flexible than the calibration measurement.A visual metrology algorithm is proposed to measure the pedestrian’s height in the uncalibrated video. Firstly, the foreground pedestrian is extracted in each frame image by simple frame difference algorithm. Then the major axis of the foreground is estimated to determine the pedestrian’s top(head) and bottom(foot) points, and two parallel virtual planes is constructed by the top/bottom points to determine the horizontal vanishing line, while the vertical vanishing point is obtained by the intersection of the major axes. Finally, the pedestrian’s height is estimated by the cross ratio theorem. This method is based on the reasonable hypothesis that a pedestrian’s height has only very small difference during the normal walking process. By using motion information sufficiently, it is shown that the visual metrology may be determined from the video, without knowledge of 3D scene information, nor of the camera’s internal calibration. Experimental results show that the average relative error is less than 2.2%, so the precision can achieve the requirement.