Research On Video Object Detection And Tracking Technology For Small-scale Flying Platform

The object detection and video object tracking are popular in the research field of computer vision.At present,these technologies have been widely used in traffic,security,military investigation and other aspects.However,when it is applied on small-scale flying platform,since the object is often far away from the platform and the target has a small proportion in the field of view,many existing algorithms are difficult to achieve good results.In addition,due to considerations such as payload and endurance,the performance of equipment which is on small-scale flying platform is generally low,many existing algorithms are difficult to transplant directly.In view of the above difficulties,this paper first analyzes the existing object detection and video object tracking algorithms,then find out several better performance algorithms as the basis for improvement.After this I checked the improvement direction through consulting the development in related research fields.Specific research content is as follows:(1)In the aspect of object detection,this paper selects the SSD object detection algorithm as the starting point for improvement,and then,referring to the recent development of compress convolutional neural network,this paper decided to use the Mobile Net V2 backbone network replace the original VGG-16,to build a light-weight SSD object detection algorithm.On this basis,this paper makes a deep analysis of the reasons for the poor performance of similar algorithms in the small object detection task.Then,inspired by the focal loss theory,I introduce the focal loss function into this algorithm,by adjusting the parameters of the focal loss function to increase the weight of the small object in training,and then improve the ability of the algorithm to detect small object,so that the whole algorithm can be better applied to small-scale flying platform.Finally,through simulation experiments and comparison with some similar light-weight object detection algorithms,it is proved that this paper proposed algorithm can significantly improve the recall rate of object detection.(2)In the aspect of video object tracking,this paper chooses the TLD algorithm as the starting point of improvement,by analyzing the weakness of the original TLD algorithm,it is decided to improve its accuracy and robustness by increasing the re-detection ability.Inspired by the visual saliency object detection technology,I decided to introduce the saliency object detection algorithm which is based on cellular automata into TLD algorithm.Since the algorithm uses local re-detection,the amount of computation in this algorithm does not increase much,but it significantly improves accuracy and robustness.Finally,the proposed algorithm in this paper is compared with the original TLD algorithm on VOT dataset,which proves the superiority of the proposed algorithm.