Active target tracking, as a core technology in computer vision and intelligent systems, is of critical significance for achieving precise target monitoring and efficient visual information acquisition. While conventional approaches primarily emphasize continuous tracking duration, they often overlook the critical aspect of visual information acquisition. This paper presents a novel active target tracking approach based on an improved Twin Delayed Deep Deterministic Policy Gradients (I-TD3) algorithm, which simultaneously achieves stable target tracking and maximizes visual information capture. Our key contributions include: (1) a novel target appearance model integrating two-dimensional information entropy with absolute observation angles; (2) an autonomous learning system incorporating prioritized experience replay, average Q-value estimation, and curriculum learning strategies; and (3) an adversarial game-theoretic reward mechanism that provides precise action feedback. Extensive experiments conducted in the Multi-Agent Particle Environments (MPE) simulation platform demonstrate the superior performance of our approach through comprehensive comparative and ablation studies.