With the rapid development of autonomous driving technology, the contradiction between the increasing processing requirements of vehicles and the resource-limited on-board processors is increasingly prominent. The emergence of vehicular edge computing solves the physical limitation of on-board resources and enhances the computing capacity of a single vehicle. However, due to the delay-sensitive of vehicular services in autonomous driving scenarios, how to choose the appropriate access technology to satisfy the delay constraint of vehicular services has become a challenge. In this paper, two kinds of V2X communication technologies, namely short range communication (DSRC) and cellular vehicular communication (C-V2X), are considered comprehensively, and a task offloading model of V2X heterogeneous vehiclular network is proposed. Firstly, the characteristics of vehicle mobility are analyzed, and the on-board resources are virtualized. Then, the task offloading problem is modeled based on the principle of semi-Markov decision processes(SMDP), and the state, action, reward and transition probability are defined respectively. Finally, the optimal task offloading strategy is obtained based on the reinforcement learning intelligent algorithm, and the performance of the algorithm is proved to be better than the greedy algorithm through a large number of numerical simulations.