Real-Time Visual Target Detection and Tracking Via Unmanned Ground Vehicle

摘要

Vision-based autonomous robot field is becoming rapidly popular, according to the Artificial Intelligence revolution. The proposed system is a composition of Unmanned Ground Vehicle vision-based target tracking robot. Target tracking is useful in multiple real-life issues such as the assistance and security fields. Acquitted visual input processing is applied through using OpenCV library. The object detection process is done based on a pre-built deep learning object detection model. YOLOv3-tiny is used to detect objects, which is a light computation-cost version comparing to the original YOLOv3 model and the other complex deep-learning networks. Object to track is specified to be a human only. The target tracking algorithm is based on a sequence of mathematical equations with Region of Interest and stream’s frame coefficients. coefficients refer to the values of locations according to x-axis and y-axis of the frame. A simple mathematical technique is used for the delayed feedback issue. Ultra-sound technique is used for collision avoidance. The locomotion of UGV is based on transmitted commands from algorithm to the motors through local network connection. The results show an autonomous behavior, streamlined and accurate locomotion of tracking.