A review on object detection algorithms based deep learning methods / Wan Xing ... [et al.]

One of the most dynamic areas in AI research is object detection, a field that continues to evolve due to advancements in chip computing power and deep learning techniques. The central goal of object detection is to identify objects and determine their precise locations by leveraging image processing technology. This application finds utility across diverse industries, such as traffic management, crime scene investigation, and assisted driving. The training process for deep learning-based object identification involves several key steps, thoroughly exploring the data preprocessing, neural network design, prediction, label allocation, and loss calculation. Deep learning-based object detection algorithms can be categorized into three main types: end-to-end algorithms, two-stage algorithms, and one-stage algorithms. Additionally, algorithms can be further divided into anchor-free and anchor-based variants, based on whether bounding boxes are predetermined. This paper begins by reviewing the history and evolution of object detection. It also outlines significant milestones for backbone networks, traditional object detection models, and deep learning-based object detection models, all according to their chronological progression. Furthermore, examples of essential performance evaluation metrics and datasets are provided, while highlighting pressing issues and emerging trends within the field that demand further investigation.