基于门控循环单元的障碍物检测方法

Obstacle Detection Method Based on Gated Recurrent Unit

针对现有障碍物检测方法存在检测精度和速度不足的问题,提出一种基于门控循环单元(Gated Recurrent Unit,GRU)的障碍物检测方法。通过构建单向GRU网络提取障碍物的几何特征,结合障碍物的时序特征实现障碍物检测。在GRU的基础上,提出三种优化模型注意力GRU模型、正则化GRU模型以及双向GRU模型用于提高障碍物检测精度或检测速度。为了验证所提方法的有效性,在真实采集的数据集上进行实验,结果表明,相较于卷积神经网络,GRU网络能够以较高的精度和速度实现障碍物的检测,其中,正则化GRU模型收敛速度更快,检测速度更高,综合性能最好。

Aiming at the problems of insufficient detection accuracy and speed of the existing obstacle detection methods,a method based on gated recurrent unit(GRU)was proposed.The geometric characteristics of obstacles were extracted by constructing a oneway GRU network for obstacle detection,and the obstacle detection was realized by combining the time-series characteristics of obstacles.Based on the GRU,three optimized models,the attention GRU model,the regularized GRU model,and the bidirectional GRU model,were proposed to improve the accuracy and speed of obstacle detection.In order to verify the effectiveness of the proposed method,experiments are performed on a real collected data set.Experimental results show that the GRU network can achieve the detection of obstacles with higher accuracy and speed compared to a convolutional neural network.Moreover,the regular GRU model has faster convergence speed,higher detection speed,and best overall performance.