Applied Research On Multiple Object Tracking In Complex Environment And Its Key Problems

Object tracking is a key research subject in the field of computer vision research,and its application is very wide and active,such as missile guidance,video surveillance,medical diagnosis,UAV systems and robot navigation.Based on image processing technology,object tracking integrates electronic optics,computer and testing technology to form a synthesis.The key issue of object tracking is that the apparent change of the tracked object can make the algorithm robust and reliable.The apparent change is mainly reflected in two aspects,one is the change against the external environment,such as light or shadow,and the other is the occlusion.There is no stable method in detecting and tracking objects in such uncontrollable environments,and due to the poor scalability in handling multiple objects,the existing methods have no good generalization ability when dealing with complex data sets,therefore,the multi-object tracking issue in these uncontrollable environments was focused in this paper.The research is summarized as follows:1)Multi-object tracking is divided into two stages: object detection and data association,in which,the object detection results have a great impact on the data association results.In order to improve the precision of the object detection,the neural network technology was adopted to enhance its stability and reliability.Although the traditional object detection algorithm appears to be efficient in algorithm complexity,it is difficult to deal with the sudden change of the environment and it is insufficient in ability to adapt to the new environment.However,the unique learning mechanism of neural network can handle this issue satisfactorily with minimal manual intervention.In the application of neural network,the traditional BP neural network and the new convolutional neural network technology were used to carry on the experiment respectively,and several enhancement measures were taken to improve the reliability of the small learning network.Experiments showed that neural network can effectively solve the issue of object tracking,and can satisfactorily deal with the changes in the environment,such as light or shadow.In addition,the controllable partial occlusion can also be detected in the object detection phase.These results have greatly enhanced the reliability of the data association stage,and can be properly used in embedded devices and other hardware limited conditions.2)On the aspect of multi-object tracking,if the traditional object detection algorithm is adopted,the detection error under the more complex situations will be expanded in the data association stage,which will affect the performance of data association algorithm.One idea at this moment is to couple these two subproblems of object detection and data association into a single and overall optimization problem,thereby making good use of the complementarities between these two sub-problems.In the selection of algorithm,an improved algorithm based on mean shift was selected for object detection,and a linear programming method was used in data association.Both of these two traditional algorithms are characterized by simplicity and efficiency,and more importantly,they can be transformed into optimized solutions which could facilitate the coupling of sub-problems.Finally,the coupling algorithm was obtained by coupling these two common algorithms,and the data set was applied to test the effect.The experimental results indicated that the coupled algorithm can maintain a high degree of accuracy when dealing with both severe displacement and object disappearance.