Visual Object Detection In Traffic Scenes Based On The Cognitive Mechanism

Object detection is one of the core tasks of biological vision and computer vision.There is a strong regularity between objects and the scene structure,e.g.,the location or size distribution of objects in scenes.Making full use of the scene structure information is expected to significantly improve the performance of object detection in complex scenes.At the same time,in the field of visual neuroscience,some studies have proved the correlation between scene structure and objects,which can accelerate visual search and improve the accuracy of object search.Considering the stable spatial structure characteristics of traffic scenes,and based on the correlation between object distribution and scene structure,this study explores the guidance or constraint of scene structure information on object detection for the object detection task in traffic scenes and establishes a more effective object detection model of traffic scenes.Therefore,firstly,based on the spatial structure information of the traffic scene,an object proposal algorithm is designed to verify the effectiveness of the traffic scene structure guiding object detection.In addition,based on the semantic segmentation,this study establishes the traffic scene structure information coding（such as vanishing point,road boundary,and other structure information）,and integrates it into the neural network for object detection,so as to improve the detection performance of the network model for objects of different categories and scales.The main research contents include the following two parts:（1）Firstly,this paper uses a vanishing point detection algorithm to extract the main structure information of traffic scene,and uses the relationship between scene structure and object distribution as a priori information to design an adaptive object box selection method.Combined with typical candidate object detection algorithms（such as Edge Boxes）,a region proposal method based on scene structure is established in this paper.The experimental results show that under the guidance of traffic scene structure,the region proposal method can achieve more accurate potential object region selection based on fewer candidate regions,which proves the important role of scene structure information in promoting object detection.（2）In addition,combined with the deep learning technology,this paper first uses a semantic segmentation network to obtain the segmentation results of the main elements of the traffic scene（such as pavement）,and establishes the representation of the spatial structure of the scene.Then,based on the object detection network,this paper establishes a multi-task network model integrating the structural relationship between the object and the scene and improves the object feature representation of the network model by predicting the relationship between the object and the scene.Experiments show that the object detection network integrating scene structure information can improve the object detection performance of different categories and scales.