To address the challenges in balancing accuracy and real-time performance for navigation path recognition methods of wheat harvesters, this study introduces a lightweight design based on the DeepLabv3+ model. Specifically, the original backbone network is replaced with MobileNetV3-Large, and the ReLU activation function is substituted with Leaky_ReLU. To further reduce computational load, the three dilated convolutions with different dilation rates within the atrous spatial pyramid pooling (ASPP) module are replaced with depthwise separable convolution (DSC) employing identical dilation rates. To ensure the generalizability of the wheat harvester navigation path recognition research, images of wheat harvest areas were captured under six typical environmental conditions: strong light, low light, back lighting, front lighting, shadows, and field edges. The road information within the collected images was annotated by the Labelme tool, thereby constructing a wheat harvest region data set. In the path extraction stage, key points were acquired from the segmentation mask maps using the horizontal scanning method, and the navigation path was fitted utilizing a piece wise cubic B-spline algorithm. Experimental results demonstrate that the improved DeepLabv3+ model achieved a segmentation accuracy of 98.04% and intersection over union (IoU) ratio of 95.20%, respectively, with a video image processing frame rate of 7.5 frames per second. The average pixel error and average distance error for navigation path recognition were 7.4 pixels and 37 mm, respectively. The wheat harvester operated at a travel speed of approximately 1.5 m/s, and the path recognition time per single frame was only 0.15 seconds. This performance effectively meets the real-time and accuracy requirements for wheat harvester operation. This research provides a theoretical foundation and technical support for enhancing the autonomous navigation capabilities of wheat harvesters.