Cassava (Manihot esculenta Crantz) is a major tuber crop worldwide, but its mechanized harvesting is inefficient. The digging–pulling cassava harvester is the primary development direction of the cassava harvester. However, the harvester clamping–pulling mechanism cannot automatically adjust its position relative to the stalks in forward movement, which results in clamping stalks with a large off-center distance difficulty, causing large harvest losses. Thus, solving the device’s clamping location problem is the key to loss reduction in the harvester. To this end, this paper proposes a real-time detection method for field stalks based on YOLOv4. First, K-means clustering is applied to improve the consistency of cassava stalk detection boxes. Next, the improved YOLOv4 network’s backbone is replaced with MobileNetV2 + CA, resulting in the KMC-YOLO network. Then, the proposed model’s validity is demonstrated using ablation studies and comparison tests. Finally, the improved network is embedded into the NVIDIA Jetson AGX Xavier, and the model is accelerated using TensorRT, before conducting field trials. The results indicate that the KMC-YOLO achieves average precision (AP) values of 98.2%, with detection speeds of 33.6 fps. The model size is reduced by 53.08% compared with the original YOLOv4 model. The detection speed after TensorRT acceleration is 39.3 fps, which is 83.64% faster than before acceleration. Field experiments show that the embedded model detects more than 95% of the time at all three harvest illumination levels. This research contributes significantly to the development of cassava harvesters with intelligent harvesting operations.