Object detection and tracking seem an important basis for the logistics industry where multiple packages are moved on conveyor belts at a time. In order to sorting, counting, and tracking parcels or packages, object detecting and tracking are prevalent and challenging tasks that the logistics industry is looking forward to accomplish. Towards providing a fully automated solution, it is necessary to address the limitation of sorting, detecting, and tracking parcels or packages from conveyor belts inside parcel distribution hubs.

Data annotation is one of the important parts of both parcel detection and tracking due to the training of the deep learning framework that uses data annotation and ground truth annotation is not 100 % accurate. The existing Ground Truth Concept is being utilized to prepare datasets where various tools enable human experts to draw bounding boxes over the objects/packages. However, ground truth data does not fulfill the robustness in terms of accuracy.

Thus, preparing a set of human-annotated unique IDs for all of the ground truth bounding boxes can overcome the limitation to train a deep learning model accurately. Addressing the aforementioned problem, we consider developing a novel application that shall enable users to assign id by clicking specified bounding boxes and stores in ground truth data (JSON files). We use OpenCV and Tkinter GUI to display the bounding box from the ground truth dataset.

We utilized performance matrices to detect objects and tracking them afterward.

Euclidean Distance: Euclidean distance (ED) or Euclidean metric is called centroid tracking and the algorithm is a combination of a process of calculation of straight-line distance between two points in euclidean space to find the distance between the current frame and the referenced frame for each object. The method relies on (i) existing object centroids and (ii) new object centroids between subsequent frames. Based on a threshold distance between objects (e.x. 50%), the model categorizes the objects as similar otherwise creates a new label.

We demonstrate our application to determine the 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.

The Hungarian Algorithm, also known as Kuhn Munkres algorithm, is an algorithm for combinatorial optimization that solves the assignment problem in polynomial time with two possible ways to formulate the problem, (i) as a matrix or (ii) as a bipartite graph. We adopt the Hungarian algorithm method to identify and track parcels due to its ability to identify whether a parcel from one frame is the same as another parcel from another frame.

We demonstrated our tools on ~6,200 datasets consisting of thousands of bounding boxes and achieved an accuracy of 99.18 \% in finding the correspondence between parcels and tracking the parcels. Such accuracy indicates the robustness of our framework. In the future, we aim to demonstrate our model with other deep learning models including neural networks to compare the accuracy with introduced applications.