Research On 3D Object Detection Technology Based On Key Points

As we all know,in the field of computer vision,object detection has always been important.With the development of deep learning,3D object detection research is also advancing rapidly.Compared with the 2D algorithm that only provides the category of the object and the length and width boundary boxes,3D object detection combines the depth information and can provide the spatial information such as the position,direction and size of the object,which is more reliable for describing scenes in the 3D real world.In 3D object detection technology,3D object detection based on key points has always been the focus of research,but it has some problems such as low detection accuracy and speed and difficult to balance.In this dissertation,we focus on the requirements of lightweight scenes.In order to improve the accuracy and speed of the 3D object detection model of key points,an in-depth study on the 3D object detection model of key points based on the depth-separable convolution and attention is carried out.In the 3D object detection model of key points,a lightweight attention module combining channel attention and spatial attention is proposed,which can effectively improve the accuracy of 3D object detection of key points.The residual block based on depthseparable convolution is used to replace the residual block based on ordinary convolution to reduce calculation cost and improve detection speed.Experiments were designed to verify that the accuracy of 3D object detection model of key points can be improved by using channel attention or spatial attention alone or by using different serial modes of channel attention and spatial attention.It is verified that the serial mode of channel attention and spatial attention is the most significant in improving the average accuracy of each index.Secondly,experiments are carried out to verify that residual blocks based on depth-separable convolution can effectively improve the speed,and the effect of accuracy is negligible.Finally,this model can bring some improvement in speed and precision and realize the balance between speed and precision.For more complex scenarios on the accuracy of the main requirements,this dissertation introduces a more complex Res Net50 as backbone network,and puts forward a key points 3D object detection model based on pyramid squeeze attention,with a lower learning attention weighting model complexity,in a more granular level of extracting multi-scale spatial information effectively.It can effectively integrate local attention and global attention to establish long-term channel dependence relationship.The experiment verifies that Res Net50 network is more suitable for the refined scene where accuracy is more important than speed,and the key point 3D object detection model based on pyramid squeeze attention can effectively improve the accuracy.A vehicle 3D detection system is designed from the perspective of application.The system uses the two models proposed in this dissertation to detect vehicle objects and verifies the validity of the proposed model from practical application.Therefore,the 3D object detection models based on key points proposed in this dissertation has certain feasibility and effectiveness,and has important research significance and value.