Research On Motion Trajectory Based On Monocular Camera And Object Detection Technology

With the continuous improvement of the level of science and technology in recent years,the technology of intelligent machine is developing rapidly.It is mainly used in robots,AR(Augmented Reality),unmanned aerial vehicles,driverless cars areas,etc.In these areas,the intelligent machine often get the perception through the visual sensor from the outside world.On the one hand the intelligent machine can carry out its own motion estimation for relative positioning by the surrounding scene;On the other hand intelligent machine detects the objects around,so the machine can make a further decision.In this paper we research the trajectory generation of monocular vision and the detection of objects in the scene.The specific research jobs are as follows:The motion of the camera is estimated by the epipolar geometry to generate the trajectory of the camera so as to achieve the relative positioning of the machine itself.The traditional motion estimation of camera is based on the five-point method to extract the camera rotation information from the essential matrix,which contains multiple solutions,and we need to eliminate the abnormal solutions artificially.To solve this problem,in this paper we proposes a new two-point method based on least squares to extract the unique solution of the rotation information of the monocular camera from the equation set of essential matrix,and it combine the camera translation in the world coordinate system to generate the trajectory incrementally,In order to solve the problem of interference of light and noise during camera motion,in this paper we uses bundle adjustment to optimize the motion of the camera and improve the performance of the anti-jamming of the algorithm.Experimental results show that the algorithm runs reliably,and the method is of higher robustness in indoor environment.SSD based on convolutional neural network is a kind of object detection model which is based on 16-layer VGG network,but SSD network structure has the problem which can't effectively detect small object.Therefore,this paper proposes an SSD network architecture which based on a 50-layer residual network.The residual network structure can make the data more freely choose the propagation path and reconstruct the learning process without changing the network complexity,so it can realize a deeper SSD network structure,and enhance the ability of data expression of the model.Experiments show that in the process of training the network model,it is easier to converge than before,and small objects can be effectively detected in the video by the network model.