摘要
针对现有行车障碍预测方法存在目标单一性、预测速度慢和准确性不佳等问题,提出一种融合空间注意力机制的卷积神经网络Coll-Net以及基于Coll-Net的车速控制和障碍方向判定策略。模拟驾驶员通过视觉信息判断障碍的机制,以单目视觉图像作为输入,首先对图像做预处理得到感兴趣区域,然后利用残差块网络提取区域内的空间特征;采用空间注意力机制对特征通道上的原始特征进行重新标定,获得通道权重;再将通道权重归一化后加权到通道对应的空间特征上,以此挑选关键特征,最后送入全连接层和Sigmoid函数中生成预测概率。行车根据障碍预测概率实时确定行车速度并根据多窗口的概率预测值判定障碍方向。实验表明,Coll-Net模型的障碍预测准确率达到96.01%,F1-score达到0.915,模型推理时间仅需24 ms,能够实时检测车辆、行人、护栏、墙体等多种障碍物,并且在低对比度光照环境下仍表现出良好的预测能力,基于Coll-Net的车速控制和障碍方向判定策略在Udacity Self-Driving数据集上表现出强有效性。
To address the limited detection targets,slow processing speed,and low accuracyof existing methods for driving obstacle prediction,this paper proposed an improved convolutional neural network called Coll-Net merged with spatial attention,a suitable speed control policy,and an obstacle direction determination method based on Coll-Net.Coll-Net imitated the vision mechanism of judging obstacles during driving,preprocessed the input monocular vision images to obtain the region of interest,and extracted the spatial features using a deep residual network framework.After collecting the spatial features,Coll-Net recalibrated the original features on the spatial feature channels by using the mechanism of spatial attention,which evaluated the features of every channel,improved the important ones,and then rescaled the weights of every channel and assigned the normalized weights to the corresponding spatial features in order to select critical features.The output feature map is connected by a fullyconnected layer;then,a normalized obstacle probability range of 0 to 1 is generated by a sigmoid function.Moreover,this paper proposes a driving policy,that controls the driving speed and predicts the obstacle direction according to the generated probability by Coll-Net.Experiment results indicate that Coll-Net prediction accuracy on standard datasets reaches 96.01%and the f1 score reaches 0.915.Coll-Net performs well in detecting diverse obstacles such as cars,pedestrians,guardrails,and walls in real time(24 ms for inference),as well as in low-contrast conditions.Moreover,the driving policy based on Coll-Net is validated using Udacity Self-Driving datasets.
作者
雷俊锋
贺睿
肖进胜
LEI Jun-feng;HE Rui;XIAO Jin-sheng(Electronic Information School, Wuhan University, Wuhan 430072, China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2020年第8期1850-1860,共11页
Optics and Precision Engineering
基金
国家自然科学基金资助项目(No.41771457)。
关键词
障碍预测
单目视觉
深度学习
注意力机制
智能驾驶
obstacles prediction
monocular vision
deep learning
attention mechanism
intelligent driving
