Research On UAV Video Image Pedesrian Tracking And Scene Understanding

Consumer small UAV has a wide range of application prospects.It is an important technical function of the UAV to track a pedestrian and scene understanding with video image sensor.The corresponding robust real-time pedestrian tracking algorithm and scene understanding algorithm are the key technologies.which has important practical application value.Aiming at the characteristics of video images taken by UAV,a robust and efficient method for real-time tracking of pedestrian targets and an algorithm for scene understanding are studied in this paper.After analyzing of the particularity of UAV video image and the reasons for the poor effect of the existing algorithm on the UAV video,this paper presents a target pedestrian tracking algorithm which is suitable for a moving platform.The algorithm is designed to follow the following principles:(1)The ACF(Aggregated Channel Features)detector is introduced to obtain the best matching accuracy between the bounding box and pedestrian real size;(2)Compared to the traditional algorithm which use a "every frame detect" strategy,the local area ACF object detect algorithm combined with median optical flow with“interval frame detect”improve the algorithm's calculation speed dramatically;(3)In order to overcome the interference of multiple pedestrians and obstacles,an online human color feature model matching with high robustness is proposed.The experimental analysis comparison between proposed pedestrian tracking algorithm and existing algorithm is conducted in this paper,the experimental results show that on the UAV video,the proposed algorithm performs better than traditional algorithm in center location error,overlap precision and distance.Outdoor scene understanding based on computer vision is an important research direction of intelligent robot technology.Detection and recognition for different kinds of object is a technical problem.In this paper,the YOLO deep neural network is used to detect and identify the target of UAV video image,so as to realize the outdoor scene understanding.At the same time the test on scene understanding algorithm and experimental analysis on UAV video is conducted.The experimental results show that the algorithm has certain effect,and is able to detect and recognize object such as pedestrian,car and building,proves that the scene understanding algorithm based on deep neural network is feasible,but still need to optimize original algorithm to improve detection ratio.