<p>This article presents a path planning algorithm for obstacle avoidance of autonomous vehicles. The proposed method utilizes a neural network-centric path sampling strategy which combines a neural network and sampling-based path planning algorithm. The neural network proposes a driving path of the ego vehicle based on the information of driving environment. The neural network has been trained based on expert demonstrations. Subsequently, path primitives are sampled around the proposal path estimated by the neural network. Applying the safety condition to the primitives, the optimal path primitive is used as the final driving path. Exploiting expert knowledge guides the path sampling process and improves sampling efficiency. The performance of the proposed algorithm has been evaluated via simulation studies and vehicle-in-the-loop tests. Evaluation results indicate that the proposed neural network-centric path sampling derives collision-free driving path by exploiting reduced computational resources.</p>