Abstract:

Convolutional neural network has achieved good detection performances on vision-based robotic grasp detection. However, most of them have too many calculation parameters and are not suitable for the system with limited resources. To solve this problem, a lightweight grasp detection regression model SqueezeNet-RM（SqueezeNet Regression Model） is proposed based on the lightweight neural network SqueezeNet and the idea of using multiple bypass connections to enhance feature reuse in DenseNet, and SqueezeNet-RM is used to extract multi-modal features（RGD） from RGB-D images to predict the best grasp pose for two-finger robotic gripper. On the standard Cornell Grasp Dataset, compared with the classical grasp detection method, the proposed lightweight grasp detection network takes up less storage space and shows faster detection speed and higher generalization performance without decreasing the detection accuracy. The storage space occupied by the proposed model is 86.97% smaller than that of AlexNet and 3 times faster, which is more feasible to deploy on FPGA（Field Programmable Gate Array） or resource restricted robotic grasp detection system.

Key words: deep learning, DenseNet, SqueezeNet, robotic grasp detection, lightweight convolutional neural network

摘要：

卷积神经网络在基于视觉的机器人抓取检测任务上取得了较好的检测效果，但是大多数方法都有太多的计算参数，不适合资源有限的系统。针对这个问题，基于SqueezeNet轻量级神经网络，结合DenseNet多旁路连接加强特征复用的思想，提出了轻量级抓取检测回归模型SqueezeNet-RM（SqueezeNet Regression Model），并使用SqueezeNet-RM从RGB-D图像中提取多模态特征，预测二指机器人夹持器的最佳抓取位姿。在标准的康奈尔抓取数据集上，提出的轻量级抓取检测网络与经典的抓取检测方法相比，在保证检测准确率不降低的情况下，模型占用更少的存储空间，表现出更快的检测速度和更高的泛化性能，所提出的模型占用的存储空间比AlexNet模型减少86.97%，平均检测速度快3倍，适用于FPGA（Field Programmable Gate Array）或者资源受限的移动机器人抓取检测系统。

关键词: 深度学习, DenseNet, SqueezeNet, 机器人抓取检测, 轻量级卷积神经网络