Since the obstacle detection methods based on color and height information could only detect some of all obstacles in farmland, a detection method based on frequency was proposed. The wavelet multi-resolution decomposition was developed to find the frequency layer of crops and it was observed that the total frequency of crops was more dominant than others and the distribution of crops row were considered. Then the positions and horizontal dimension of possible obstacles crossed the crop rows in the image were detected based on frequency distribution of the selection layer. The others could be detected in the highest frequency layer due to the lower frequency of this kind of obstacles. Then the stereo rectification and the prior frequency knowledge of obstacles were adopted to confirm if the detection was obstacle. The experiment showed that the proposed method could detect the mound, the edge of farmland and other obstacles effectively. The average time of processing each frame was 79ms.