In vehicular edge computing (VEC), few existing works focus on co-scheduling both on intra-vehicle and inter-vehicle for deep neural network (DNN) inference. Moreover, all of them ignore the selfishness of vehicles and load balance among vehicles, which results in a lack of guarantee in quality of inference services for a long time. We aim to fill this gap by investigating the co-scheduling both on intra-vehicle and inter-vehicle with consideration of selfishness and load balance for vehicles in VEC. Specifically, we formulate a co-scheduling problem with objective of maximizing total revenue of all vehicles, under the constraints of tolerant response time of tasks, tolerant energy consumption of vehicles, etc. To address the formulated problem, we propose an incentive algorithm based on coalition game, to encourage vehicles with underload to share their resources for alleviating the load of vehicles with overload. Meanwhile, the proposed algorithm schedules DNN inference tasks both on intra-vehicle and inter-vehicle to make full utilization of resources in VEC. The proposed algorithm is evaluated on a platform based on OpenStreetMap and SUMO, to simulate real vehicular scenarios. Simulation results on the platform show that, our proposed algorithm outperforms the state-of-the-art for all cases, in terms of system revenue.