Real time target tracking system based on deep learning

The real-time visual object tracking system covers the detection and identifies the target in a video sequence or in fixed images, with a boundary box (x and y coordinates) and keep tracking of that object in ambiguous environments like illumination and none stationary cameras, without a human intervention. Even more than that, we need a real time object tracker, either by increasing the hardware specifications, which are limited sometimes, or by improving practical algorithms, to meet the requirements. In this study we present an analysis and knowledge research for deep learning and tracking techniques, then a crossbred tracking method is proposed for creating a real-time and accurate deep learning based tracking system. Our Detect Track (DETRAC) framework combined of You Only Look Once object detector (YOLO) with Kernelized Correlation Filter (KCF) tracker, to achieve a high accuracy, precision, robust correction and target relocation structure. In order to increase the tracking capability and the overall performance, we used YOLOv3 architecture, and we provided a specialized custom dataset of quadrotors (drones) to train our network. The combination precision was 80%, which is 8% higher than YOLO alone, and 58% higher than KCF alone, and runs at 31 frame per second on our experimental hardware.