Research On Dynamic Object Situation Awareness For Autonomous Driving

Situation awareness of dynamic objects is the core task for the environment perception system of autonomous driving,and it is an important part of scene understanding and modeling.However,the situational awareness technology for dynamic objects is still not mature,which is one of the bottlenecks restricting the development of self-driving cars.Focused on the issue of“Situation Awareness of Dynamic Objects for Autonomous Driving”,this dissertation fully explores three key problems of moving object detection,tracking,and trajectory prediction by contextual clues mining and multi-source information integration,aiming at enhancing the perception of dynamic objects and understanding of dynamic scenes for self-driving cars.The main contributions and innovations of the dissertation are summarized as follows:Aiming at the perception of moving objects in dynamic scenes,we propose a novel bio-inspired 3D moving object detection method based on Li DAR points.Inspired by the motion detection mechanism of compound eye organisms ——Elementary Motion Detector,this approach develops two kinds of bionic visual neural loops with different connection methods,and achieves efficient motion detection in a coarse-to-fine fashion.In order to achieve further cognition of the moving point clouds,a lightweight cascaded3 D object recognition network is designed to accurately estimate the bounding box and category of the moving object.The most significant feature of this approach is that the detector can detect the specific movements by changing the connection between the receptors according to different specific demands,thereby improving the calculation efficiency.We create a 3D moving object detection dataset based on the KITTI benchmark,and conduct qualitative and quantitative experiments.Compared with the start-of-the-art methods,the performance of the proposed approach has been greatly improved.In addition,experiments on self-driving vehicles further verify the effectiveness and robustness of the approach.Aiming at the problem of expressing the temporal relationship of moving objects,a3 D moving object tracking method based on multi-modal data fusion is proposed.Multisensor calibration is the basis for multi-modal data fusion.In order to get rid of the need for specific markers,we propose an automatic extrinsic calibration algorithm for cameraLi DAR in natural scenes.Using the ego-motion information and the corresponding reflection of the same feature points in different modal data,the algorithm realizes automatic calibration of external parameters.On this basis,we propose a camera-Li DAR fusion multi-object tracking method.The algorithm takes advantage of multi-modal data to accurately detect all objects and extract the apparent features.The motion prediction model performs 3D motion inference to overcome occlusion problems.Based on the similarity score of the association matrix,the weighted Hungarian algorithm is used to achieve object association.Experiments on the KITTI MOT task and self-driving vehicle verify the effectiveness of the proposed method.Aiming at the problem of understanding the intention of moving objects,a hierarchical multiple contextual clues integrated trajectory prediction method is proposed.Specifically,given the continuous states of moving objects and high-definition maps,an LSTM-based encoder extracts the motion features to express the driving characteristics of each actor.Meanwhile,an attention-based graph module is applied to extract the interaction feature of moving objects and accurately model the binary relationship of actors.The scene features are extracted from high-definition vector maps by convolution neural networks.Combining these three types of attribute features,the decoder module then infers the future trajectory.This method fully excavates the context information and comprehensively considers the motion characteristics,interaction characteristics,and environmental constraints of moving objects.Thus it can more accurately predict the motion trajectory and behavior intention of moving objects.We evaluate the proposed approach on three widely-used datasets,and state-of-the-art results demonstrate the effectiveness of our approach.Especially in the Apollo Scape trajectory prediction task,the proposed method ranks third place in comprehensive task and first place in pedestrian trajectory prediction.