Object Detection and Tracking: Deep Learning-based Framework with Euclidean Distance, IoU, and Hungarian Algorithm
Disciplines
Computer and Systems Architecture | Robotics
Abstract (300 words maximum)
Object tracking is an important basis for the logistics industry where multiple packages are moved on conveyor belts at a time. Accurate datasets and efficient benchmarks are a few of the several problems for both object detection and tracking for training the deep learning-based framework. Preparing 100% accurate correspondence between objects throughout different frames by assigning human annotated unique_attributes to train framework efficiently over ground truth data. In this research, we develop an (i) OpenCV-based framework that allows the user to assign human-annotated identification between objects and (ii) a novel application for object detection and tracking. We utilize the assigned attributes to train the deep learning model accurately and adopt various evaluation parameters including euclidean distance, intersection over union (IoU), and scale-invariant feature transform (SIFT) to measure the accuracy of an object correspondence or tracking. We also adopt the Hungarian algorithm to increase the efficiency in determining correspondences between objects and apply our framework to human-annotated ground truth datasets comprising ~1,000 images and the same amount of JSON files. Our demonstration achieved 94.53 % accuracy in object detection, finding correspondence, and object tracking. In future studies, we are aiming to apply a neural network to draw a comparison of identified accuracy.
Academic department under which the project should be listed
Department of Computer Science
Primary Investigator (PI) Name
Hossain Shahriar
Additional Faculty
Maria Valero, CCSE-Information Technology, mvalero2@kennesaw.edu
Object Detection and Tracking: Deep Learning-based Framework with Euclidean Distance, IoU, and Hungarian Algorithm
Object tracking is an important basis for the logistics industry where multiple packages are moved on conveyor belts at a time. Accurate datasets and efficient benchmarks are a few of the several problems for both object detection and tracking for training the deep learning-based framework. Preparing 100% accurate correspondence between objects throughout different frames by assigning human annotated unique_attributes to train framework efficiently over ground truth data. In this research, we develop an (i) OpenCV-based framework that allows the user to assign human-annotated identification between objects and (ii) a novel application for object detection and tracking. We utilize the assigned attributes to train the deep learning model accurately and adopt various evaluation parameters including euclidean distance, intersection over union (IoU), and scale-invariant feature transform (SIFT) to measure the accuracy of an object correspondence or tracking. We also adopt the Hungarian algorithm to increase the efficiency in determining correspondences between objects and apply our framework to human-annotated ground truth datasets comprising ~1,000 images and the same amount of JSON files. Our demonstration achieved 94.53 % accuracy in object detection, finding correspondence, and object tracking. In future studies, we are aiming to apply a neural network to draw a comparison of identified accuracy.