The intelligence of agricultural machinery is the hotspot of current agricultural intelligent research, and the visionbased environmentaware technology is the key technology to realize the intelligence of agricultural machinery. An algorithm based on wavelet transform was proposed to extract the illumination invariant features of farmland images. According to the Retinex illumination model, the image included two parts as the illumination component and the object reflection component. The illumination component can be regarded as the lowpass filtered image of the original image, that was, the lowfrequency part of the original image. Therefore, by removing certain low frequency components in the original image, it was possible to obtain the illumination invariant feature. The original farmland image was preprocessed, including clipping and normalization. The preprocessed image was multilevel decomposed by Haar wavelet base to obtain the high and low frequency components of the image. The highfrequency coefficients after wavelet decomposition were updated by the threshold method, and the multiscale reflection model was reconstructed to extract the illumination invariant features. Finally, the experimental study on illumination invariant feature extraction and crop route acquisition was carried out. The result proved that the feature image extracted by the proposed algorithm was little affected by natural illumination and can retain the object features in the scene to a great extent. At the same time, crop route extraction had high precision under different illumination conditions, and the route error was within ±2°, which can meet the accuracy requirements of agricultural machinery navigation. In addition, on NVIDIAs Jetson TX2 hardware platform, the proposed algorithm took less than 300ms, and the cameras forwardlooking distance can reach 20m, which can meet the realtime requirements of the normal operation of agricultural machinery.