自动驾驶环境感知多任务去耦-融合算法

Decoupling-fusing algorithm for multiple tasks with autonomous driving environment perception

自动驾驶车辆在行驶过程中,需要对行人和车辆同时完成目标检测、实例分割和目标跟踪三个任务。提出一种基于深度学习的环境感知模型同时对三个任务进行多任务学习。首先,通过卷积神经网络对连续帧图像提取时空特征;然后,通过注意力机制对时空特征进行去耦再融合,充分利用任务间的相关性,实现不同任务对时空特征的差异化选择;最后,为平衡不同任务间的学习速率,使用动态加权平均的方式对模型进行训练。在KITTI数据集上的实验结果表明,所提模型在目标检测方面,比CenterTrack模型F1得分提高了0.6个百分点;在目标跟踪方面,比TraDeS(Track to Detect and Segment)模型多目标跟踪精度(MOTA)提高了0.7个百分点;在实例分割方面,比SOLOv2(Segmenting Objects by LOcations version 2)模型AP_(50)和AP_(75)分别提高了7.4和3.9个百分点。

In the process of driving,autonomous vehicles need to complete target detection,instance segmentation and target tracking for pedestrians and vehicles at the same time.An environment perception model was proposed based on deep learning for multi-task learning of these three tasks simultaneously.Firstly,spatio-temporal features were extracted from continuous frame images by Convolutional Neural Network(CNN).Then,the spatio-temporal features were decoupled and refused by attention mechanism,and differential selection of spatio-temporal features was achieved by making full use of the correlation between tasks.Finally,in order to balance the learning rates between different tasks,the model was trained by dynamic weighted average method.The proposed model was validated on KITTI dataset,and the experimental results show that the F1 score is increased by 0.6 percentage points in target detection compared with CenterTrack model,the Multiple Object Tracking Accuracy(MOTA)is increased by 0.7 percentage points in target tracking compared with TraDeS(Track to Detect and Segment)model,and the AP_(50)and AP_(75)are increased by 7.4 and 3.9 percentage points respectively in instance segmentation compared with SOLOv2(Segmenting Objects by LOcations version 2)model.