针对现有的目标检测算法部署在机械臂上会占用大量系统资源、检测实时性差、模型参数量大等问题,提出一种改进YOLOv5的目标检测算法.首先,在YOLOv5骨干网络融入Shffle-netV2模块代替原本的焦点模块和跨级局部暗格网络,实现网络的轻量化;其次,主干网络末端颈部将带有残差精化的上下文转换器注意力(context-transformer with residual refinement,CTR3)模块嵌入到所设计的骨干网络中,来改善模型的特征提取潜力并减小采样带来的损失;再次,头部采用基于一致性的自适应信息传递注意力(similarity-based adaptive message passing attention,SimAM)模块来增强特征的跨尺度融合能力;最后,为提高检测算法对目标的边界框回归速率和样本稳固性,引入新型加权交并比(weighted intersection over union,WIoU)函数,在目标检测网络设计完成后,将其部署在机械臂上并完成验证.实验结果表明:文中改进的检测算法准确率为96.5%;检测速度为每秒82帧,相比原算法提高32帧;参数量为430×104,相比原算法减少了39%;每秒浮点数计算次数为6.7次,约为原计算次数的1/3.数据表明改进后的检测算法检测速度快、参数量少、占用内存小,满足精准检测的前提下提高了检测效率.
Robotic arm target detection algorithm combined with deep learning
Existing target detection algorithms deployed on robotic arms would occupy a large a-mount of system resources,have poor real-time detection performance,and have a large number of model parameters.To address these problems,an improved YOLOv5 target detection algorithm was proposed.First,the ShfflenetV2 module was integrated into the YOLOv5 backbone network to replace the original focus module and cross-level local dark grid network to achieve lightweight network.Secondly,the end neck of the backbone network would have a context converter atten-tion module with residual refinement(CTR3)module embedded into the designed backbone net-work to improve the feature extraction potential of the model and reduce the loss caused by sam-pling.Then,the head used a similarity-based adaptive message passing attention(SimAM)module to enhance the cross-scale fusion ability of features.Finally,in order to improve the bounding box regression rate and sample stability of the detection algorithm for the target,a new weighted in-tersection over union(WIoU)function was introduced;after the target detection network design was completed,it was deployed on the robotic arm and verified.Experimental results show that the accuracy of the improved detection algorithm in this article is 96.5%;the detection speed is 82 frames per second,which is 32 frames higher than the original algorithm;parameters is 430×104,which is 39%lower than the original algorithm;and the number of floating point calcula-tions is 6.7 times per second,which is about 1/3 of the original number of calculations.The data shows that the improved detection algorithm has fast detection speed,small number of parame-ters,and small memory usage,improving detection efficiency while meeting the premise of accu-rate detection.
robotic armobject detectionYOLOv5attentionloss function
万方数据
维普
