Modern vehicles face cyber threats due to Internet connectivity and insecure in-vehicle networks such as CAN (Controller Area Network), which lack encryption and authentication functions. Existing study proposes a machine learning-based intrusion detection systems (IDSs), but the methods require high computational resources and cannot follow dynamic change of the network condition. Therefore, this study proposes a lightweight, real-time intrusion detection system designed for in-vehicle networks with limited computational resources. Our system employs an autoencoder-based anomaly detection model on edge devices and leverages a server-based retraining mechanism to adapt to changing of traffic patterns. In the proposed method, traffic on the CAN bus is processed in 3-s time windows, and the statistical features are extracted and used as the input of the machine learning model. The anomaly is detected based on the difference between the input and output of the model. We implement the proposed system using the network simulator of CAN (ICSim) and an embedded device (Raspberry Pi 4), demonstrating that our approach can be realized in real-time deployment even under resource-constrained environments.