| Environment perception is a hot and challenging research issue in the field of Autonomous Land Vehicle(ALV)as it perceives environment information and acts as the corner stone that supports all complementary components of the ALV.It directly manifests the capability of autonomous driving and the level of intelligence.In relevant studies of the environment per-ception in the unstructured environment,accurate and robust road area detection and obstacle detection are the crux for autonomous driving.The research work focusing on the two aforementioned issues is performed and presented in this dissertation:To tackle the problem of road area detection,this work has proposed two algorithms in terms of feature window design and road area detection framework,namely,a feature window segmentation method,and a road area detection method based on multi-scale superpixel.Towards the problem of negative obstacle detection,a negative obstacle detection method based on multiple LiDARs and compositional features is proposed.Concretely,the research content and contributions of this dissertation are as follows:(1)Due to drastic illumination variations,disturbance of vegetation,variety of road material and fuzzy road boundary,the content and texture between different regions can be het-erogonous where traditional superpixel segmentation algorithms fail to achieve desirable performance in feature window partition.In this dissertation,a super pixel segmenta?tion algorithm based on local competition mechanism is proposed.Local competition mechanism is exploited to construct energy term and optimize the whole energy func?tion,the algorithm is self-adaptive to the content and texture of the image,thus improves the robustness in segmenting different regions in the image.(2)In order to solve the problem that the appearance features of similar target are prone to in-consistentence in different superpixel feature windows,a superpixel oriented feature ex?traction method is proposed.By analyzing and comparing the classification performance of different color and texture features of superpixels,a multi appearance combination feature which is applicable for unstructured environment is proposed.Then,a road de?tection method based on multi-scale superpixel under the framework of CRF is proposed to solve t!he problem of road area detection in the unstructured environment.Firstly,it performs multi-scale segmentations to obtain superpixels;Secondly,classifies superpix-els with the XGBOOST classifier,outputs the road probability map;and then calculates the energy term with road probability map,which is later assigned with weights defined by the similarity between pixel and its corresponding superpixel;lastly,aggregates the weighted energy term under multiple scales as input of conditional random fields to seg-ment the final road region.Experimental results in the unstructured environment show that the proposed method achieved decent road area detection performance.(3)For the negative obstacle detection problem in unstructured environment,a multiple L-iDAR and combination features based approach is proposed.Due to the limited area coverage and the insufficient perception resolution of a single LiDAR,a multiple Li-DAR installation method with complementary capability is designed.According to the installation mode of LiDAR,a negative obstacle detection method based on combina-tion features is proposed.The method does not rest on the ground plane assumption and the height difference between point cloud data.Instead,it detects the negative obsta-cles through the local geometric characteristics and the distribution characteristics of the point cloud.In the spirit of spatiotemporal fusion,the negative obstacle feature points from multiple frames captured by multiple sensors are fused by the Bayes rule.Lastly,DBSCAN is used to cluster the feature points,and the results are rasterized to extract the negative obstacle area.Experimental results show that the proposed method produce good performance in detecting negative obstacle in unstructured environment. |
