As the density of people flows in large-scale exhibitions and public venues continues to rise, traditional intelligent crowd perception and crowd control methods have problems such as low accuracy, slow response, and unreasonable path planning in terms of dynamic adaptability, real-time decision-making, and refined scheduling. To this end, this paper introduces the fusion technology of artificial intelligence and the Internet of Things, and proposes a crowd control modeling and optimization framework based on convolutional neural network (CNN) driven. This paper collects dynamic feature information such as crowd density, speed, direction vector, etc. in real time through multi-source heterogeneous perception data such as camera visual stream, WiFi/Bluetooth trajectory and infrared counter, and constructs a spatial topological model of the exhibition scene; CNN is used to extract features and predict congestion risks of multi-dimensional crowd dynamic data, and at the same time combines multi-objective optimization strategies (such as Pareto frontier analysis and reinforcement learning scheduling) to achieve optimal evacuation path generation and dynamic guidance instruction issuance, and fine-grained control of crowd flow rate, distribution balance and evacuation efficiency. Experimental results show that the proposed method can improve the congestion prediction accuracy by 7.8%, shorten the average evacuation time by 10.4%, and improve traffic efficiency by 12.1% in a typical exhibition scenario, effectively achieving the performance optimization of intelligent perception of crowd flow and dynamic control of details.