ANALYSIS OF CLASS-AGNOSTIC SINGLE-OBJECT TRACKING METHODS

Abstract

This study presents a comprehensive analysis of four class-agnostic single-object tracking algorithms: KCF (Kernelized Correlation Filter), CSRT (Channel and Spatial Reliability Tracking), SAMURAI, and MMTrack. The research evaluates their performance across multiple criteria including processing speed, localization accuracy (measured by LaSOT AUC), robustness to occlusions, illumination changes, and scale variations. The experimental results demonstrate distinct performance profiles for each method: KCF achieves the highest processing speed (201 fps on CPU) but shows limited accuracy (22% LaSOT AUC) and poor resilience to occlusions and scale changes; CSRT provides a balanced trade-off between speed (80 fps) and accuracy (28% AUC) with improved robustness to partial occlusions and lighting variations; SAMURAI, built upon SAM2 with motion-aware memory mechanisms, delivers exceptional accuracy (70-74% AUC) and excellent robustness to various challenging conditions, but requires substantial computational resources (0.4 fps on CPU, 13 fps on GPU); MMTrack implements a unified token-based approach for vision-language tracking, achieving comparable accuracy (70% AUC) with moderate processing speed (4 fps CPU, 54 fps GPU) and superior adaptability to scale changes. The analysis confirms that no universal solution dominates across all scenarios, and the optimal choice depends on specific application requirements, available computational resources, and performance priorities. The study establishes a methodological framework for informed algorithm selection in video surveillance, autonomous systems, and robotics applications.