In vehicular ad hoc networks (VANETs), video communication makes a significant contribution to quality of experience (QoE) for people on the road. However, the selection of the video source is an impediment to video delivery due to the high mobility and dynamic topology of VANETs. An improper provider not only leads to frequent interruptions of communications, but the transmission of the invalid video fragments would also result in the waster of precious bandwidth. To address the issue, a novel video source decision scheme, named Cluster and Dynamic Overlay based video delivery over VANETs (CDOV), is proposed in this paper. By the on-demand clustering approach, nodes with the same video requirement/supply and moving features are clustered. Further, in a cluster, an overlay tree is constructed dynamically based on the relation between supply and demand, in which all requesters can find their greedy optimal source easily. In addition, the intracluster communication and head-RSU communication are designed for video streaming over this network structure. Using extensive simulations, the effectiveness of the proposed scheme is demonstrated. Compared with two existing works, the proposed solution is capable of obtaining lower startup latency and higher delivery ratio.