Optimizing the Visual Perception Algorithm for Dynamic Obstacles in Complex Environments for Weeding Robots

Visual perception of dynamic obstacles in the complex environment of hilly orchards suffers from poor detection performance, real-time performance, and processing efficiency. This paper proposes an optimization scheme for the dynamic obstacle perception algorithm based on the fusion of multi-scale spatiotemporal features. It builds an improved YOLOv7-CSPDarknet53 backbone network, introducing dynamic convolution kernels to adaptively adjust the receptive field. It also incorporates an optical flow-depth information fusion module, utilizing PWC-Net to calculate inter-frame optical flow and fuse it with depth data acquired by an RGB-D camera to construct a four-dimensional spatiotemporal feature tensor for improved moving object segmentation accuracy. This paper also develops a trajectory prediction model, estimating obstacle motion parameters through a state-space model to achieve trajectory prediction. The algorithm achieves an F1 score of 0.91 in mixed lighting scenarios, with a stable latency of 78-92ms and a high frame rate. This improves the robustness and real-time performance of dynamic obstacle detection, providing reliable technical support for autonomous navigation of weeding robots in unstructured environments.