Multi-task network based on hierarchical aggregation and height semantic information perception
Aiming at the problem of current multi-task networks lack analysis of relationships of multiple tasks and inherent characteristics of urban scene images,a multi-task network based on hierarchical aggregation and height semantic information perception is proposed. Firstly,in order to enhance the capacity of the feature extraction network,hierarchical aggregation module is used to learn the interdependence between multi-layer features. Through sharing and independant design,the fusion of shallow layer features and deep layer features is achieved to feed the required characteristics for different downstream tasks. Secondly,there is a certain degree of height difference in the road scene image,and the pixel-level distribution between the horizontal segmentation parts is significantly different. Height-aware module is used to introduce the prior information. The structure is simple and efficient. The results show that the proposed method outperforms similar methods in various performance metrics on BDD100K,and at the same time,lane line datasets TuSimple and CULane are re-labelled and expanded for multi-tasks test,achieves higher precision than existing methods,validating the effectiveness of the proposed algorithm.
万方数据
维普
