基于卷积神经网络的仓储物体检测算法研究

Research on Warehouse Object Detection Algorithm Based on Convolutional Neural Network

针对仓储环境中物体检测公开数据集匮乏的问题,通过摄像机采集真实仓储环境中包含货物、托盘和叉车的大量图像进行标注,创建了一个仓储物体数据集.同时针对传统物体检测算法在仓储环境中检测准确率较低的问题,将基于卷积神经网络的DSOD应用于仓储环境中,通过在自己创建的仓储物体数据集上从零开始训练DSOD模型,实现了仓储物体的准确性检测.该算法的mAP达到了93.81%,比Faster R-CNN、SSD分别提高了0.04%、1.44%;并且模型大小仅有51.3 MB,比Faster R-CNN、SSD分别减小了184.5 MB、43.4 MB.实验结果表明,该算法获得了较为满意的仓储物体检测效果,其在仓储物体检测领域具有一定的实用价值.

Considering the lack of public datasets for object detection based on the warehouse environment,a large number of images containing cargos,trays and forklifts in real warehouse environment are collected and labeled to build the warehouse object dataset. Meanwhile,aiming at the problem that the traditional object detection algorithm has lower detection accuracy in warehouse environment,the deeply supervised object detectors(DSOD)based on convolutional neural network is applied to the warehouse environment,and the DSOD model is trained from scratch on the self-built warehouse object dataset,and the accuracy detection of the warehouse object is realized. The mean Average Precision(mAP)of this algorithm reaches 93.81%,which is higher than that of Faster R-CNN and SSD by 0.04 and 1.44 points respectively,and the model size of this algorithm is only 51.3 MB,which is lower than that of Faster R-CNN and SSD by 184.5 MB and 43.4 MB respectively. The experimental results show that the algorithm has a relatively satisfying warehouse object detection effect,and it has certain practical values in the field of warehouse object detection.