摘要
无人驾驶垃圾清扫车通常通过路沿检测实现自主路径选择.传统深度学习方法虽然具有较好识别效果,但计算量过大不符合实际应用需求.为此,提出一种基于YOLOv5轻量化的改进GDFE-YOLO算法.采用Ghostconv和C3Ghost对原网络模型主干进行替换;再引入可变形卷积替代Neck部分中的传统Conv卷积模块;最后将损失函数替换为Focal-EIOU Loss.实验结果表明,GDFE-YOLO算法在参数量、计算量和模型大小上分别较原模型降低了16.2%、15%和19.4%,检测速度提高12%;路沿识别均值精度为96.1%.GDFE-YOLO相较于YOLOv5s算法的识别精度仅下降0.9%,同时基于整体路沿的线性拟合检测策略,其单点的精度下降影响较小,因此所提出算法能够实现清扫车的轻量化路沿检测要求.
Abstract
Driverless garbage sweepers often achieve autonomous path selection through curb detection.Although the traditional deep learning method has a good recognition effect,the amount of computa-tion is too large and does not meet the practical application requirements.Therefore,an improved GDFE-YOLO algorithm based on YOLOv5 lightweight was proposed.Ghostconv and C3Ghost are used to replace the backbone of the original network model,and then the deformable convolution DC-Nv2 is introduced to replace the traditional Conv convolution module in the Neck part,and finally the loss function is replaced by Focal-EIOU Loss.Experimental results show that the GDFE-YOLO algo-rithm reduces the number of parameters,the amount of computation and the size of the model by 16.2%,15%and 19.4%,respectively,and the detection speed is increased by 12%,and the average accuracy of curb recognition is 96.1%.Compared with the YOLOv5s algorithm,the recognition accura-cy of GDFE-YOLO is only reduced by 0.9%,and based on the linear fitting detection strategy of the overall curb,the accuracy of a single point has little impact,so the proposed algorithm can realize the lightweight curb detection requirements of sweepers.
NETL
NSTL
万方数据