With the growth of urban Intelligent Transportation Systems (ITS), vehicular tasks increasingly overload roadside edge nodes (RSUs), challenging low-latency and energy-efficient service. We design a cloud–edge–end architecture where Unmanned Aerial Vehicles (UAVs) act as mobile computing resources to assist task offloading. The task offloading problem is formulated as a Markov Decision Process (MDP). We propose a hierarchical algorithm, HADDQN, which combines Dueling DQN for learning optimal offloading policies and the Hungarian Algorithm (HA) for resource scheduling and UAV–Smart Lamp Post (SLP) matching. UAV lifecycle states (idle, active, parking, charging, sleep) are incorporated to improve realism and energy efficiency. Simulations in urban traffic scenarios show that HADDQN achieves lower latency, reduced energy consumption, and better UAV trajectory optimization than baselines.