基于卷积神经网络的多模型交通场景识别研究

Research on Multi-model Traffic Scene Recognition Based on Convolution Neural Network

利用人工智能中的视觉分析技术,实现对高分辨率交通视频中出现的各个目标类别进行实时目标检测、语义分割和目标追踪。数据集结合BDD100K和Mapillary Vistas。训练中不仅对模型中的参数进行调整,还对多个模型进行改进与创新。目标检测模型使用EfficientNet-B1作为主干网络,使用ASPP与改进后的FPN作为脖颈网络,通过引入多种模型训练技巧,对模型进行优化,最终结果减少约2.3倍的参数量,在不同数据集上的准确率都有所提升。目标追踪使用DeepSort追踪算法对多个目标类别进行追踪计数。语义分割使用Encoder-Decoder结构,使用EfficientNet-B4作为主干网络,参照U-Net++网络使用卷积层作为特征提取模块,反卷积层作为上采样模块,通过联结不同大小的特征图,得到最终输出结果。将改进语义分割模型与MobileNetV2和DeeplabV3网络结合的模型进行对比,减少约1.35倍的参数量。实验证明,通过深度学习算法提取鲁棒性特征能够为自动驾驶和辅助驾驶场景中的检测识别提供便利。

The visual analysis technology in artificial intelligence is used to realize real-time object detection,semantic segmentation and object tracking for each object category in high-resolution traffic video.The dataset combines BDD100K and Mapillary Vistas.In the training,not only the parameters in the model are adjusted,but also several models are improved and innovated.The object detection model uses EfficientNet-B1 as the backbone network and uses ASPP and improved FPN as the neck network.By introducing a variety of model training skills,the model is optimized.The final result reduces the number of parameters by about 2.3 times and improves the accuracy on different datasets.Object tracking uses the DeepSort tracking algorithm to track and count multiple object categories.The semantic segmentation uses the Encoder-Decoder structure and EfficientNet-B4 as the backbone network,and referring to the U-Net++network,it uses the convolution layer as the feature extraction module and the deconvolution layer as the up sampling module,and obtains the final output result by connecting the feature maps of different sizes.The improved semantic segmentation model is compared with the model combined with MobileNetV2 and DeeplabV3 network,and the number of parameters is reduced by about 1.35 times.Experiments show that extracting robust features through deep learning algorithm can facilitate the detection and recognition in automatic driving and assisted driving scenes.