Reasearch On Pedestrain Detection And Tracking In Video

Pedestrian detection and tracking based on video is the direction of a research hotspot in computer vision, the study has a broad application prospect in the field of intelligent monitoring. Intelligent surveillance system not only has important practical applications, but also plays an important role to artificial intelligence, pattern recognition, and other areas of computer vision. The tasks of pedestrian object detection and tracking are used to detect pedestrian objects in the video, and track the pedestrian objects in real-time. After long time of unremitting research, although many research scholars in the field have put forward many good theories and algorithms of pedestrian detection and tracking, but practices show that the pedestrian detection and tracking algorithms have more or less shortcomings. Considering the actual application scenarios and environment, so proposing an intelligent pedestrian detection and tracking system which can be applied to a variety of complex situations becomes the current urgent task. This paper studies the pedestrian object detection and tracking in key technical issues. The main contributions of this paper can be concluded as follow:(1) Use the mixtures of multi-scale deformable part model to characterize the pedestrian model. The model can adapt to the changeable appearance of the non-rigid pedestrian, and increase the pedestrian detection accuracy. The model is based on DPM (deformable part model) and composed of many DPM. DPM uses a star-structured part-based model defined by a root filter plus a set of parts filters and associated deformation models. Experiments show that the correct detection based on the pedestrian detection algorithm of the mixed multi-scale deformable part model is significantly higher than the traditional detection algorithm.(2) Use the fast feature pyramids based on the prediction algorithm to calculate the pedestrian characteristics. The traditional extraction of pedestrian features needs to calculate the each scale of the feature pyramids, the fast feature pyramids adopted in this paper just calculate a scale in every octave, and other scale features can be got by using the prediction algorithm directly to compute. So that this algorithm greatly reduces the calculation and saves the calculation of computing resampling image. Experiments show that the time of using the fast algorithm to extract the pedestrian characteristics is less than the traditional algorithm, and no loss detection performance.(3) Use the particle filter based on a time-varying state space model. In this paper, join the pedestrian movement acceleration to the time-varying state space model, the model can overcomes the blindness of the spread of particles and raises the use efficiency of particle set in the process of pedestrian tracking. The model changes with the change trend of target movement speed, and more close to the pedestrian's actual movement situation, so that the model improves the effectiveness and guidance of the spread of particles. Experiments show that the algorithm improved can accurately locate the location of the pedestrian when pedestrian motion is non-uniform.(4) Use the observed model based on color gradient direction histogram. The observation model is combined with the gradient intensity and color value in the gradient direction to model the target feature. The observation model contains the color information of the traditional observation mode, but also contains the pedestrian space information and texture information. So it can describe the characteristics of the pedestrian more accurately, and make the pedestrian matching between the video frames is more accurate. Experiments show that the modified tracking algorithm can tracks pedestrians more accurately than the traditional algorithm.