Autonomous vehicles are required to operate among highly mixed traffic during their early market-introduction phase, solely relying on local sensory with limited range. Comprehending and navigating complex urban environments is potentially not feasible with sufficient reliability using the aforesaid approach. To address this challenge, our present work aims to conceptualize, implement and evaluate an end-to-end cooperative perception system using novel techniques. A comprehensive yet extensible modeling approach for dynamic traffic scenes is proposed first. It is based on probabilistic entity-relationship models, accounts for uncertain observations and combines low-level attributes with high-level relational knowledge in a generic way. Second, the design of a holistic, distributed software system based on edge computing principles is proposed as a foundation for multi-vehicle high-level sensor fusion. In contrast to most existing approaches, this solution is designed to rely on Cellular-V2X communication and employs geographically distributed computation nodes as central data brokers. The modular proof-of-concept implementation is evaluated in different simulated scenarios to assess the system's performance both qualitatively and quantitatively.