关键词: 抓取检测, 注意力机制, SqueezeNet, 单目标物体检测, 深度学习

Abstract:

In order to solve the problem of single-object grasping detection of robotic arms， a SqueezeNet model algorithm based on attention mechanism （CBAM） was proposed. Firstly， the deep vision grasping system was described， and the hand-eye calibration of the grasping system was completed. The data set was preprocessed by randomly clipping， flipping， adjusting contrast， and increasing noise， effectively expanding the object capture data set. Secondly， a lightweight SqueezeNet model was introduced. It uses a five-parameter method to characterize the two-dimensional grab frame， which can complete the target capture without increasing the difficulty of network design. Thirdly， a plug-and-play network with an attention mechanism was introduced to weight the incoming feature maps in the channel and spatial dimensions. The SqueezeNet model-grabbing network was optimized and improved. Finally， the improved CBAM-SqueezeNet algorithm was verified on the public data sets Cornell grasping dataset and Jacquard dataset. The grab detection accuracy is 94.8% and 96.4%， accuracy increased 2% than the SqueezeNet network. The CBAM-SqueezeNet network grabbing method has a reasoning speed of 15 ms， which balances grabbing accuracy and running speed. The paper conducted experiments on the Kinova and SIASUN arm， and the success rate of network capture was 93%， which was faster and more efficient.

Key words: grab detection, attention mechanism, squeezenet, single-target object detection, deep learning