YOLOv11-Based Split Large-Kernel MBConv with Channel-Gated Slimmed Attention: A HarmonyOS-Embedded Framework for Efficient Crowd MOT

With the development of economy, population migration from rural areas to urban centers has intensified, imposing significant pressure on statistical indicators for social order maintenance and security risk assessment due to increased urban population density. Addressing the computational constraints and precision balance in real-time multi-object tracking for embedded device-based crowd behavior analysis, this study proposes YOLO-MGNet - a lightweight neural network architecture based on the YOLOv11 framework. The architecture incorporates a Split Large-Kernel MBConv module that decomposes traditional large-kernel convolution into depthwise separable convolution and adaptive receptive field expansion units. An innovative Channel-Gated Slimmed Attention mechanism is introduced to dynamically prune channels and allocate gated weights, effectively addressing small-target and occlusion challenges. The system employs cascade feature matching for human target tracking. Compared with the original YOLOv11, YOLO-MGNet achieves improvements of 4.26% and 3.41% in precision and recall rates respectively, with mAP@0.5 increased by 9.72% while reducing parameters by 15%. Implemented on RK3588 devices running HarmonyOS, the system achieves 98% accuracy with efficient crowd tracking performance on embedded platforms.