Lane detection plays a vital role in the research of CWS (Collision Warning System), which is based on the radar system that detects positions and velocities of vehicles over the detectable range. In CWS, the first target of warning is the vehicle in the same lane as the vehicle with CWS. Lane detection is required to distinguish this vehicle among whole vehicles detected by radar. It is especially necessary in the curved lane, because the radar beam crosses over the other lanes. For this purpose, the machine vision technique is considered a powerful method. In this article, the image-processing algorithm for recognizing the lane curve direction, before locating the lane, is proposed. By using near and middle field lane image, it can recognize the lane curve direction in the far field which contains imperfections. In this algorithm, lane features like positions and angles of lane marks in the image frame are extracted first. Then, using the neural network of inputs from lane features, the lane curve direction is determined. Test results for 2,000 frames of real road images showed a success recognition rate of over 90%.