LSARE-Net: Target-Guided Augmentation and Lightweight Edge Detection for Real-Time Behavior Monitoring of Captive Juvenile Sika Deer

Abstract

Real-time machine-vision monitoring of group-housed juvenile sika deer is challenging because targets are often small, densely distributed and partially occluded, while edge deployment requires compact models. This study developed a lightweight system for non-contact behavior recognition under commercial farm-pen conditions. A fixed-camera dataset was constructed with five annotated categories: resting, standing, feeding, social and abnormal. The dataset was limited to one month of recordings from one fixed camera monitoring six juvenile sika deer in a single pen at one farm in Jilin Province, China. To reduce background interference during training, a Multi-target Bionic Foveal Augmentation (MBFA) strategy was proposed to generate target-guided, spatially nonuniform degradation based on the distribution of deer instances, without introducing additional inference cost. A Lightweight Selective Alignment Reorganization and Efficient Detection Network (LSARE-Net) was then designed to enhance local behavioral cues, contextual feature representation, and cross-scale fusion. On the test set, LSARE-Net achieved 91.26% recall, 93.47% mAP@0.5 and 80.24% mAP@0.5:0.95 with 0.70 M parameters and a 1.7 MB model size. After TensorRT conversion, it reached 62.64 FPS on an NVIDIA Jetson AGX Orin. These results demonstrate the feasibility of compact frame-level behavior detection under the evaluated pen conditions. The accompanying prototype organizes detections into behavioral summaries and abnormal-cue records for manual review.