Research On Pedestrian Head Pose Estimating Method Based On Surveillant Video

With the development of computer vision,visual perception has been widely applied in various industrial scenarios,such as intelligent transportation system,security and protection system,etcetera.In these scenarios,head pose estimation of pedestrian in surveillant videos has been a researching hotspot.Still and all there are many problems to be tackled in pedestrian detection,pose estimation and pedestrian tracking,which mainly reflect in the low resolution of pedestrian in image,the top view of the surveillant camera,the change of illumination,shadows and occlusions and so on.Aiming at the theme of head pose estimation of pedestrian in surveillant images,this thesis focuses on several pivotal tasks which are highly related to pedestrian headpose estimation based on surveillant images.The major studies of the thesis are outlined in the followings:1.Setting up the remoting surveillant system for acquiring video data,designing a simplifying algorithm for making data set.(1)A surveillant camera is mounted in the high place of a building located in a block where jaywalking occurs frequently.The camera acquires images of different time periods and diverse weathers.The images are captured by the surveillant camera from birds'-eye-view and remoting distance,which include various challenges of deficient resolution,shadow,illumination variation and occlusions.(2)Thirteen experimenters are requested to simulate the scenario of jaywalking,meanwhile a IMU mounted on their heads are recording the information of their head pose.(3)The data are then made into pedestrian detection data set.An auxiliary method based on motion analysis techniques is proposed to simplify the labeling work.2.A two-step method for pedestrian head detection based on deep learning is proposed,which applies to surveillant images and other remotely captured images.Firstly,a pedestrian detector is trained to locating the position of pedestrians in the image.Secondly,the images of detected pedestrian region are extracted.Then,intersection over union(IOU)is utilized to divide the image into positive and negative samples.Lastly,a head detector is trained to extract the bounding-box of heads accurately based on the pedestrian images.The detecting framework leveraged in both pedestrian and head detectors is YOLOv3 due to its outstanding performance.Experiments show that the method is effective for accurately locating head region from remote surveillant images.3.A head orientation recognizing method applying to low-resolution images based on Convolutional Neural Networks(CNNs)is proposed.In consideration of the deficiency of resolution of the heads images,head orientation recognizing task is defined as a 10-category classifying mission for the yaw angle of the heads in the project.Firstly,the two-step head detecting method is utilized to extract head region from source images.Afterwards,the IOU criterion is leveraged to select valid head images.Finally,the pose estimating samples composed of the head images and their category labels are input into the network constructed from ResNet-50 for training the classifying model.A serial of experiments are enforced to show the appealing performance.4.An algorithm for pedestrian head pose measurement in surveillant image sequence is proposed.The algorithm is mainly composed of multi-pedestrian tracking module,surveillant image calibration module and head pose measurement module.Firstly,the multi-pedestrian tracking module uses YOLOv3 as a pedestrian detector.Then,a pre-training deep neural network is used to describe the CNN feature of pedestrians.Finally,an online pedestrian tracker is implemented based on Deep SORT framework.On this basis,the head center of multiple pedestrians in each frame of the image sequence is detected.Firstly,the foreground pixels in the head image are obtained by threshold processing.Then,the vertical direction and neck position of pedestrians are analyzed by multi-direction projection,and the precise head area is extracted.Finally,the positions of the center of the head is obtained by calculating the geometric moment of the head region.Experiments show that the algorithm achieves appealing performance of pedestrian tracking with robustness to light changes,shadows and obstacles,and can effectively measure the physical location of pedestrian head center.