en.Wedoany.com Reported - A research team led by Professor Zhang Man from the College of Information and Electrical Engineering at China Agricultural University (CAU) published a study in the Chinese Academy of Engineering journal *Engineering*, titled "Multimodal Feature Representation Mechanism for 3D Detection of Agricultural Obstacles with Few or Zero Samples." This research addresses the need for safe perception of field obstacles in autonomous navigation of intelligent agricultural machinery, proposing a multimodal feature representation mechanism-based method for few-shot and zero-shot 3D detection of agricultural obstacles. It offers new insights for safe and reliable autonomous operation of agricultural machinery in unstructured farmland environments. Founded by the Chinese Academy of Engineering in 2015, *Engineering* is indexed in SCI, EI, Scopus, and others, with a latest SCI impact factor of 12.2, once ranking first among global comprehensive engineering journals.

Autonomous navigation of agricultural machinery is a key support for intelligent agricultural equipment, and safe perception of obstacles in complex farmland environments is crucial for ensuring reliable autonomous operation. In recent years, deep learning methods integrating camera and LiDAR data have made significant progress in 3D obstacle detection, but these methods typically rely on large-scale annotated training data. Farmland scenes are characterized by unstructured layouts, significant seasonal variations, and complex obstacle types, leading to high costs for multimodal data collection and annotation. The generalization ability of models under few-shot or even zero-shot conditions remains a major challenge for practical deployment.

The proposed multimodal feature representation mechanism integrates an image and point cloud pose corrector, utilizing pose information from the BeiDou Navigation Satellite System and inertial measurement units to correct sample pose deviations caused by rugged field terrain, thereby improving the accuracy, reliability, and consistency of multimodal data. Additionally, it constructs a semantic feature encoder, a geometric-intensity feature encoder, and a bird's-eye view spatial fusion decoder, unifying image semantic information with point cloud geometric and intensity data into a semantic-geometric-intensity fusion representation space. This captures key relationships between categories, enhancing the model's ability to recognize new obstacle categories under limited annotation conditions.

The team conducted field experiments at CAU's Zhuozhou Experimental Station, covering typical agricultural machinery operation scenarios such as cement roads, uncultivated land, and wheat fields, collecting multimodal data of typical obstacles including harvesters, tractors, and personnel. Results show that the proposed method achieves a good balance among detection performance, operational efficiency, and data dependency, reducing the model's reliance on training samples by 30%-40%. Under full training settings, precision, recall, F1 score, and detection speed reached 95.03%, 97.01%, 96.01%, and 16.56 FPS, respectively. In zero-shot scenarios, where obstacle categories have no corresponding training samples, the method still achieved an F1 score of 81.63%.

This achievement helps reduce the dependency on large-scale annotated data for 3D obstacle detection in complex agricultural environments, enhancing the safe perception capabilities of intelligent agricultural machinery under conditions of unknown obstacles, complex terrain, and multi-category targets. It provides technical support for autonomous navigation, dynamic obstacle avoidance, and reliable operation of smart agricultural equipment.

The study was completed in collaboration with China Agricultural University, Beijing Forestry University, and CRRC Research Institute, among others. China Agricultural University is the first affiliated institution, Professor Zhang Man is the corresponding author, and Wang Tianhai, a 2021 master's student and recipient of the 2024 university-level Outstanding Master's Thesis Award, is the first author. The research was supported by the National Key Research and Development Program of China (2022YFD2001600-2022YFD2001601). Invited by *Engineering*, Professor Zhang Man participated in the Engineering Lecture Hall's special session on "Agricultural Sensors" on June 30, 2026, presenting this research online.