The cooperation of roadside and onboard sensor data can effectively enhance the perception ability of connected and automated vehicles, especially in complex or large blind area traffic scenes. However, existing cooperative perception methods are mainly designed based on simulation environments or tested on offline datasets, ignoring the limitation of transmission bandwidth and time delay in real-world applications. In this paper, we proposed a practical vehicle-infrastructure cooperative perception framework named VICooper for autonomous driving. Lightweight sensing models are adopted for 3D object detection in roadside and onboard perception tasks. Besides, lane detection and the road map are utilized for filtering out the participants outside the drivable region in the vehicle and infrastructure coordinate system, respectively. In addition, assisted with the real-time pose of the connected vehicle, an efficient cooperation strategy is presented to achieve optimal matching for the multi-source perception data. Experiment results on the DAIR-V2X-C dataset show that the proposed method performs well in terms of improving object detection accuracy and expanding the sensing range of connected vehicles. Furthermore, real-world field tests also demonstrated the practicability and effectiveness of the proposal in real V2X applications.