The SI engine load torque is key information for many engine and power train control systems. Since the torque is not measured in production vehicles, it needs to be estimated on-line. The paper presents design and analysis of second-order and third-order Luenberger load torque estimators. With the aim to reduce the estimator noise sensitivity without deteriorating its transient performance, an adaptive Kalman filter is proposed and compared with the Luenberger estimator. The adaptation mechanism is based on a load torque change detection algorithm. The estimators are examined by computer simulations and experiments.