A Survey of Object Detection and Classification in Video Using Deep Learning Approaches

Source abstract

Object detection and classification in videos has emerged as a core challenge in computer vision, with applications spanning autonomous driving, surveillance, healthcare, robotics, and human– computer interaction. Unlike still-image object detection, video-based detection requires handling additional complexities such as temporal coherence, motion blur, occlusion, and real-time constraints. Traditional approaches relying on handcrafted features and background subtraction have given way to deep learning-based methods, particularly convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformers, and one-stage detectors such as the You Only Look Once (YOLO) family. This paper provides a comprehensive review of object detection in videos, surveying the evolution of methods from early handcrafted approaches to recent deep architectures. We examine popular models, datasets, and benchmarks, with an emphasis on accuracy–speed trade-offs and the unique challenges of real-world deployment. A comparative analysis of state-of-the-art methods highlights trends and gaps, including the struggle with small object detection, domain adaptation, and efficient training on large-scale datasets. Finally, we outline promising future directions such as transformer-based detection, self-supervised learning, multimodal video understanding, and edge-device optimization. This survey aims to serve as both an introductory resource and a roadmap for advancing video-based object detection research.