Overhead ground wire detection under complex field environment is one of the key technologies of automatic obstacle surmounting for a high voltage transmission line inspection robot. The varying illumination and wire surface condition are the key factors affecting the detection accuracy. To address the problem, a field wire detection algorithm based on off-line learning is proposed. In the training phase, the adaptive homomorphic filter is applied to the input samples to compensate illumination, followed by local binary pattern histogram feature extraction and support vector machine training to get a binary classifier. In the on-line detection phase, each sample is divided into patches, followed by classification through the trained classifier to get the wire patch candidates. Then, the random sample consensus algorithm is adopted to remove mistakenly identified patches, and the remaining candidates are fitted into a line to get the wire parameters in the image coordinate system. The results of a number of experiments with field surroundings and different wires show that the proposed method has good adaptability to varying illumination and can detect both old and new wire accurately. Furthermore, this work has laid a solid foundation for the subsequent three dimensional positioning and grasping control of the ground wire.