A Robust Real-Time Detecting and Tracking Framework for Multiple Kinds of Unmarked Object.

Researchers

Journal

Modalities

Models

Abstract

A rodent real-time tracking framework is proposed to automatically detect and track multi-objects in real time and output the coordinates of each object, which combines deep learning (YOLO v3: You Only Look Once, v3), the Kalman Filter, improved Hungarian algorithm, and the nine-point position correction algorithm. A model of a Rat-YOLO is trained in our experiment. The Kalman Filter model is established in an acceleration model to predict the position of the rat in the next frame. The predicted data is used to fill the losing position of rats if the Rat-YOLO doesn’t work in the current frame, and to associate the ID between the last frame and current frame. The Hungarian assigned algorithm is used to show the relationship between the objects of the last frame and the objects of the current frame and match the ID of the objects. The nine-point position correction algorithm is presented to adjust the correctness of the Rat-YOLO result and the predicted results. As the training of deep learning needs more datasets than our experiment, and it is time-consuming to process manual marking, automatic software for generating labeled datasets is proposed under a fixed scene and the labeled datasets are manually verified in term of their correctness. Besides this, in an off-line experiment, a mask is presented to remove the highlight. In this experiment, we select the 500 frames of the data as the training datasets and label these images with the automatic label generating software. A video (of 2892 frames) is tested by the trained Rat model and the accuracy of detecting all the three rats is around 72.545%, however, the Rat-YOLO combining the Kalman Filter and nine-point position correction arithmetic improved the accuracy to 95.194%.