A study and development of and tracking approaches in uav videos

Abstract

An Unmanned Aerial Vehicle (UAV) is an unpiloted aircraft capable of reaching areas that are not easily accessible by human beings and obtains high-quality images or videos at a low cost. It created a significant changeover in surveillance,land surveying, media, agriculture, emergency management, etc. The free-flying ability provides a perspective that no other technology can provide. The role of UAVs in surveillance and investigation in recent days demands accurate object detection and tracking. The free-flying ability results in various object detection and tracking challenges. In recent years, vision-based object detection and tracking methodologies have changed tremendously from traditional to deep learning methodologies. The capability of deep learning methodologies is to learn features automatically, and highly informative content obtained from UAVs provides more scope for exploration. The increasing deployment of UAVs in real-time scenarios,on the one hand, and the growth in vision-based object detection and tracking methodologies, on the other hand, necessitates better object detection and tracking in UAV videos. This thesis presents efficient approaches for object detection and tracking in UAV videos. A novel UAV video optimization framework, is developed for object tracking. Keyframe extraction, principal keyframe selection and object path analysis are developed to obtain optimized UAV video with negligible path deviation. The average video duration of 20.37 seconds is reduced to 1.93seconds saving 90.68% of viewing time. Further, Region-based Convolutional Neural Network (R-CNN) / Faster R-CNN along with a pre-trained VGG16 network, are used for object detection. Kalman Filter is used in object tracking due to its accurate prediction in various applications. newline