Weighing equipment has been widely used to estimate residual feed in beef cattle breeding in precision agriculture. However, manual assessment cannot fully meet the nutritional requirements of large-scale beef cattle production in recent years, due to strong subjectivity and quantification difficulties. Conventional feed monitoring techniques are also confined to the complex and costly deployment on the devices. In this study, a non-contact estimation was proposed for the residual feed volume in beef cattle feedlots using YOLOv8-SPD and projected grid integration. A precise and highly adaptable system was also developed to accurately quantify leftover feed using advanced machine vision. An improved YOLOv8 network was constructed for target detection and segmentation. The Space-to-Depth Convolution module was introduced to significantly enhance the feature extraction. The high-precision semantic segmentation of forage regions was achieved to remove the image motion blur and complex background interference under vehicle-mounted environments. Subsequently, a depth camera was utilized to acquire point cloud data. A registration strategy was employed with single-frame Iterative Closest Point for seamless multi-view point cloud stitching. Furthermore, a ground plane fitting with layered filtering, multiple Random Sample Consensus iterations, and statistical fusion was proposed to minimize the impact of ground undulations and noise on the volume calculation baseline. Finally, the feed pile volume was accurately calculated using a projected grid finite element analysis. The optimal voxel and grid dimensions were determined after parameter optimization. A series of experiments was conducted to validate the effectiveness and superiority of the model under various complex scenarios. The experimental results demonstrated that the improved YOLOv8-SPD model was achieved in the superior performance for the forage detection and segmentation tasks, where the Precision and Recall further increased to 92.16% and 93.66%, respectively. The better performance of the detection model was also achieved particularly in challenging scenarios with severe occlusion, varying illumination conditions, and motion blur caused by the moving platform. The Space-to-Depth Convolution module was effectively integrated to preserve fine-grained spatial information, which was directly contributed to highly accurate segmentation masks that isolated the feed points from complex background elements. Ground plane fitting exhibited exceptional robustness in the three-dimensional processing stage. The average distance error of the ground plane fitting was controlled to 5 mm, indicating the high precision. A reliable and stable measurement effectively prevented the reference plane from artificial elevation by residual debris or sensor noise. Furthermore, the highly accurate calculation was verified on the feeding volume using the projected grid finite element analysis. Four types of feed morphology—namely mixed forage, alfalfa, wheat straw, and rapeseed straw—were selected under diverse feed scenarios. A volume accuracy exceeded 87.91% after measurement. Specifically, the higher accuracy of 91.37% reached for the mixed forage, due to its uniform density and compact surface structure. Internal voids and surface scattering were minimized during depth sensing, whereas the irregular porosity in straw materials was introduced slightly with the manageable variations. To sum up, the non-contact estimation of residual feed volume demonstrated low deployment cost, high precision, and strong robustness against complex environmental disturbances, even under complex agricultural scenarios. The outstanding performance and reliability were realized in the non-contact estimation, compared with conventional contact weighing equipment. This finding can provide a highly precise data basis for the reliable decision-making on refined feeding, individual feed efficiency evaluation, and breeding cost optimization in modern beef cattle farms.