Analysis Of Target Motion And Its Application In Video Multi-Target Tracking

Video multi-target tracking,as an important research direction in the field of computer vision,has played a vital role in intelligent transportation,driverless and other directions.The main purpose of video multi-target tracking technology is to find the target of interest in a given image sequence or each frame of the video,and then associate the same target according to the characteristics of the target.Until now,the theory of video multi-target tracking has matured,but it still cannot achieve good tracking results,because there are many factors that affect the tracking effect,such as: occlusion or deformation of the target,multiple targets with similar appearance characteristics Complex pedestrian interaction.Accurate motion features can largely deal with the problems of appearance features,but it is difficult to extract accurate motion features in complex motion scenes.At present,most multi-target tracking methods focus on improving the visual feature extraction algorithm to extract excellent visual features from the target detection response.In the motion feature extraction algorithms,a simple linear motion model or a motion that cannot cope with complex pedestrian interactions is used.model.This article focuses on the improvement of motion feature extraction algorithms,and is dedicated to handling complex abstract and difficult to model pedestrian interactions.By learning the current mainstream methods,this paper is dedicated to extracting accurate motion features in complex sports scenes,and mainly completes the following tasks:1)The algorithm principle and training process of the recurrent neural network are studied,and several popular trajectory prediction methods are analyzed.The performance limitations of these methods are obtained.2)Aiming at the complex abstract and difficult to model pedestrian interactions in complex scenes,drawing on the pedestrian's visual system mechanism,a trajectory prediction method based on the pedestrian sight range was proposed.Compared with the current trajectory prediction methods,the trajectory prediction model can learn more For reasonable pedestrian interaction.3)In the pedestrian interaction model,fully consider the relative speed,relative movement direction and relative position relationship between pedestrians,and finally use the multilayer perceptron network to replace the traditional regression method to predict the displacement.4)Aiming at the problem that the trajectory prediction method based on recurrent neural network can only predict trajectories of a fixed length,a multi-mode trajectory prediction model is proposed,which can adaptively match the trajectory prediction model according to the input trajectory length,and build a multi-target tracking based on this The algorithm achieved better tracking performance than traditional tracking methods.By comparing the trajectory prediction error comparison experiment and the tracking index comparison experiment with the existing methods,the trajectory prediction method based on the line-of-sight range in this paper can handle complex pedestrian interactions better,and the tracking model based on this method has achieved better tracking performance.