野外环境下基于改进 Faster R-CNN的无人车障碍物检测方法

针对无人车在越野环境中障碍物检测存在特征提取能力不足和检测准确率低等问题，提出一种基于改进型Faster R-CNN卷积神经网络模型的障碍物检测方法。通过构建FPN与ResNet50组合的网络结构来实现对野外障碍物的特征提取，有效解决了特征提取时障碍物细节特征丢失和尺度变换大的问题。使用Soft-NMS代替NMS，避免了NMS非极大值抑制由于阈值难调整带来的误删除和误检问题。在每个卷积层残差块最后嵌入注意力机制，有助于特征图中有效特征信息筛选和减小计算量。试验结果表明，构建的改进型Faster R-CNN卷积神经网络模型可准确识别野外环境中的障碍物，从而验证了该模型有良好的检测能力，对提升无人车的野外感知能力具有重要意义。

Improved Faster R-CNN Technique for Obstacle Detection in Unmanned Vehicles in Field Environments

To address the challenges of insufficient feature extraction capabilities and low accuracy in obstacle detection for unmanned vehicles in off-road environments, an improved Faster R-CNN convolutional neural network model is proposed for this task. Firstly, the paper combines the FPN with the ResNet50 network for feature extraction, effectively solving the problems of losing detailed obstacle features and large-scale transformations during the process. Secondly, Soft-NMS is used instead of NMS to avoid the false detection arising from threshold adjustment difficulties in conventional non-maximum suppression. The attention mechanism is embedded at the end of the residual block in each convolutional layer, which is helpful to filter the effective features in the feature map and reduce computational demands. Finally, experimental results show that the improved Faster R-CNN model significantly outperforms other methods, which proves its superior detection capabilities. This study is important for improving the off-road perception capabilities of unmanned vehicles.