The integration of vehicle-to-everything (V2X) communication paradigms with sixth-generation (6G) wireless networks and artificial intelligence (AI) frameworks enables ultra-reliable, low-latency communication, which is essential for real-time decision-making in autonomous vehicles (AVs) and smart cities. Proprioceptive and exteroceptive sensors allow AVs to perceive both their internal states and external surroundings, ensuring rapid responses to critical events. Integrated sensing and communication (ISAC) enhances this capability by jointly leveraging perception and communication, enabling V2X systems to adapt intelligently to real-time emergencies. In this paper, we propose a probabilistic, data-driven, hierarchical, interactive, and explainable approach for an intelligent agent, i.e., a base station (BS), to learn the dynamic environmental perception from the 3D LiDAR point clouds and the strength of radio-frequency (RF) power signals between the connected BS and vehicles. An interactive coupled Markov jump particle filter (IC-MJPF) is proposed in the inference phase to leverage the probabilistic information provided by an interactive coupled generalized dynamic Bayesian network (IC-GDBN) to predict various types of LiDAR and RF power blockages, as well as to detect real-time abnormalities in an unsupervised manner arising from dynamic environmental changes. Experimental results demonstrate that the proposed approach consistently outperforms existing baseline studies, achieving superior performance in terms of blockage detection accuracy within 50 milliseconds across various blockage situations. These findings underscore the robustness and effectiveness of the proposed framework in addressing both physical and digital blockage challenges within the ISAC domain for connected V2X networks.