Automatic object tracking in optimized UAV video

Abstract

Target detection and tracking in unmanned aerial vehicle (UAV) surveillance enhance the monitoring capacity. In surveillance, the collection of data for long duration necessitates data optimization. In this paper, UAV video optimization is performed by discarding the redundant frames and preserving the object trajectory across the optimized video. The average UAV video duration is reduced from 20.37 to 1.93 s. Further on, the objects are detected and tracked automatically in the optimized video. Detection and tracking were performed with region-based convolutional neural network (R-CNN) and Faster R-CNN on Kalman filter. The experiment is analyzed with two datasets based on training and detection time, detecting the CARs at various views and performance metrics. Both R-CNN and Faster R-CNN have promising detection with various views of the CAR such as front view, back view, side view and top view. Apart from training time, Faster R-CNN holds better than R-CNN in both detection time and performance metric.