To address the challenges of insufficient wide-area information perception, poor robustness in multi-target tracking, and low accuracy in dangerous behavior recognition during major public security tasks, this paper integrates key technologies such as wide-area high-definition multi-source image stitching, deep learning-based image recognition, automatic multi-target tracking and early warning, and posture-behavior recognition of dangerous targets. A multi-modal panoramic imaging collaborative optimization framework is proposed, and a highly integrated wide-area surveillance system is developed. First, a requirement model for wide-area surveillance tasks is established using the AHP（Analytic Hierarchy Process）, based on which the overall system architecture, hardware architecture, and software algorithm architecture are optimized. Second, 360° wide-area surveillance is achieved through heterogeneous sensor collaborative perception and panoramic stitching, while deep learning networks realize multi-type target recognition with a success rate exceeding 90% in self-tested environments. Third, a multi-feature target tracking algorithm incorporating spatial attention feature extraction networks and subspace reconstruction mechanisms is designed, achieving a multi-target tracking success rate of 93.98% in self-tested environments. Further, a temporal-spatial dual-stream action recognition model（TSAR）is constructed, yielding dangerous behavior recognition accuracy exceeding 90% in self-tested environments. Experimental results demonstrate that this system effectively supports wide-area surveillance, target tracking/locking, and intelligent early warning in public activity scenarios, providing reliable solutions for complex dynamic security tasks.