Less-activated visual networks to enhance friendly following ability of lightweight robots

Jianxin Xue, Husheng Chen, Yaohua Hu, Sicheng Hua, Minyu Chen, Ling-I Wu, Xi Chang

Abstract: Lightweight robots engaged in real-time following tasks require low-latency object detection to provide effective companionship services. However, these robots are often CPU-edge devices with limited memory, leading to frequent memory access in detection architectures like the YOLO series, causing significant delays. To address this challenge, we introduce LALCNet, a less-activated visual network derived from the CPU convolutional neural network PP-LCNet. LALCNet removes unnecessary activation functions and incorporates dual reduction strategies to optimize detection speed, while minimizing accuracy loss. By replacing the backbone network in the YOLO series, LALCNet processes entire images more efficiently in memory-constrained environments. Our comparative experiments show that LALCNet boosts detection speed on CPU-edge devices by 64%, 2.66X, and 1.93X compared to YOLOv5, YOLOv8, and YOLOv10, respectively. Despite these significant speed improvements, the decline in mean average precision (mAP) at a threshold of 0.5 is minimal. LALCNet emphasizes speed-first optimization while maintaining an acceptable level of accuracy, making it ideal for real-time robot following tasks. It enhances the user experience by enabling faster, more responsive interactions without compromising detection quality.

External IDs:dblp:journals/ese/XueCHHCWC25