A deep reinforcement learning (DRL)-based method was proposed for task offloading and resource management optimization in vehicular network in order to enhance real-time performance and computational efficiency in autonomous driving task processing. A mobile edge computing (MEC) system was targeted. Roadside units (RSUs) were configured as computing and caching nodes in order to achieve dynamic task offloading and content caching optimization. Vehicle task demand variations were predicted by an improved Hawkes model, and corresponding caching strategies were generated. The task offloading and resource allocation problem was formulated as a 0-1 optimization model and converted into a Markov decision process (MDP). An improved proximal policy optimization algorithm was applied to obtain the solution. The simulation results showed that average computation delay was significantly reduced and request response rate was increased in highly dynamic vehicular environments. A 15.1% improvement in total reward was achieved compared with the baseline. The effectiveness of the method in optimizing real-time computation and caching management of vehicular network in complex traffic scenarios was validated.