Comparative Study of Deep Learning Algorithms for Counting Unique Objects

Abstract

Unique Object Counting in Images and Video Streams: A Computer Vision Problem with Broad Applications. Unique object counting in images and video streams is a basic problem in computer vision with numerous applications in intelligent surveillance, traffic analysis, crowd analysis, retail analytics, and smart city infrastructure. Conventional methods of object counting are prone to failures in complex real-world scenarios because of occlusions, scale changes, high object density, and dynamic environments. Recent breakthroughs in deep learning have greatly improved object detection, tracking, and density estimation, leading to more accurate and robust object counting algorithms.
This paper conducts a thorough comparative study of the latest deep learning algorithms for unique object counting, including YOLO (v5, v8, v9), Faster R-CNN, SSD, Mask R-CNN, YOLO with DeepSORT, FairMOT, and CSRNet. The compared algorithms are tested for accuracy in object counting, uniqueness preservation, computational complexity, and real-time processing suitability. The results of the experimental study show that tracking-based methods, specifically FairMOT and YOLO with DeepSORT, perform better than detection-based and density estimation-based algorithms in preserving uniqueness and avoiding double counting in video streams.

Keywords