Autonomous Target Detection and Tracking for Remotely operated Weapon Stations

Abstract

This Master s thesis proposes a novel implementation of an autonomous tracker in Python, which combines a deep learning detection module and a point based tracking module. An accurate detection will introduce latency if the video capture rate exceeds the processing rate. The use of a frame buffer, a key element of the combination design, will compensate for this weakness. All frames periodically skipped by the detector will be stored, and a fast tracker will process the buffer to provide an updated object prediction for the current frame. The system implementation is developed with focus on future deployment on a Nvidia Jetson TX2 embedded platform, and utilizes Google s TensorFlow object detection API and the OpenCV object tracking API. The autonomous tracker is evaluated on a number of relevant videos, with a hybrid measure combining the bounding box overlap and a new proposed distance error score. The final system configuration, with a lightweight neural network for detection and the median flow algorithm for tracking, show real-time performance on a quad-core CPU.