Research On 3D Target Detection Based On Multi-source Data Fusion

At present,the development of intelligent transportation has become a national strategy and has become an important promoter of national strategic transformation.Target detection technology,as the basic supporting technology of traffic intelligence,realizes accurate classification and positioning of perceived targets.The quality of detection performance directly affects subsequent integration decision-making control links,which is a prerequisite for a large number of intelligent advanced tasks.Object detection technology has important research significance and application value in the field of intelligent transportation.Aiming at the limitations of target detection technology and the disadvantage of visual sensors being vulnerable to environmental influences,this paper proposes a three-dimensional target detection based on laser radar and camera for intelligent driving environment,using the complementary characteristics of multiple sensors and the powerful feature extraction performance of neural networks.technology.The three-dimensional object detection algorithm is improved as follows from two aspects:depth feature information representation and multi-source data fusion:(1)For the feature extraction stage of depth feature information expression,a residual network with topology optimization is proposed,and the network weight reduction is realized by convolution kernel transformation.Under the principle of constant complexity and constant output dimensions,the optimal topology optimization model is determined through comparative experiments,and the target detection accuracy rate is improved by 4.6% compared with traditional algorithms.The convolution kernel transform aims at the size of the convolution kernel in the residual structure,and proposes the idea of sub-channel convolution first filtering and then optimization.After the comparison of the complexity,the model complexity is reduced by 30% and the detection speed reaches 53 FPS.(2)For the regional nomination phase of deep feature information expression,a regional nomination network with integrated multi-threshold is proposed.Through experiment,the influence of the parameter setting on the regression performance of the network is analyzed.It is found that the single threshold cannot satisfy the optimal regression problem for all anchor frames.The regression effect is best only when the input value is near the threshold.Therefore,referring to the integrated learning ideas in machine learning,considering the limitation of algorithm complexity,designing a regional nomination network with two thresholds,the target detection accuracy is increased by about 2%.(3)In terms of multi-source information fusion,based on image and point cloud data,a multi-source data deep fusion network is proposed.Combined with the deep fusion algorithm to optimize the 3D target detection algorithm to obtain the target'sdirection and depth and other information,to achieve a 3D description of the target.All optimization stages are experimentally verified based on the KITTI open source data set.The proposed 3D object detection algorithm achieves the lowest detection error rate in the optimization process.