A Review of Deep Learning Object Detection Algorithms

Abstract

Object detection and multi-object tracking are core foundational tasks in computer vision, serving as crucial supporting technologies for high-end applications such as autonomous driving, intelligent security, industrial inspection, and robot navigation. This paper systematically reviews the complete evolution of object detection and multi-object tracking technologies from traditional manual feature extraction methods to the era of deep learning, focusing on the core principles, technological advantages, and application limitations of single-stage and two-stage detectors. Building upon this foundation, it further summarizes the three widely accepted mainstream paradigms in multi-object tracking — separate detection and feature extraction, joint detection and feature extraction, and joint detection and tracking — and discusses in depth the key technologies under each paradigm, including motion feature modeling, deep appearance feature learning, attention mechanism optimization, and the fusion application of the Transformer architecture. Finally, it summarizes the core challenges currently facing the field (balancing real-time performance and accuracy, small object detection, occlusion robustness, etc.) and looks forward to key future development directions, providing a systematic reference for theoretical research and engineering implementation in related fields.