| With the rapid development of society and economy,security issues have become increasingly prominent.The advent of artificial intelligence has enabled rapid development of robot technology and its application in the field of security robots.Currently,robots used in security work places,especially undercarriage security robots,still rely on vision for the identification of dangerous goods.However,the complexity of the image background limits the accuracy and speed of identification.Therefore,this paper uses a multi-channel information fusion method.Identify dangerous targets to complement the lack of vision.The main research contents are as follows:1.According to the types and identification requirements of potentially dangerous goods on the bottom of the vehicle,image sensors,gas detectors and radiation sensors were selected to supplement the image system on the bottom of the vehicle,the overall hardware was adjusted,and system software development was carried out.Aiming at the fusion problem of heterogeneous sensors,the different algorithm structures,levels and types of multi-channel information fusion are analyzed in detail.Two algorithms,BP neural network fusion algorithm with simple structure and DS evidence theory fusion algorithm based on cloud model,are selected for improvement research..The image information,CH4 volume fraction,and radiation intensity were used as the input of the neural network,and the dangerous goods classification was output by the algorithm as the judgment result.2.BP neural network combining additional momentum and adaptive rate.Aiming at the problem of slow convergence speed of BP neural network,simulation experiments were performed on several improved algorithms.Finally,a combination of introducing additional momentum and using adaptive rate was selected.In accordance with the requirements of the vehicle bottom dangerous goods identification model in this paper,the parameters of the structure and the number of layers of the neural network were designed.The algorithm is used to simulate the sample.The experiments show that the improved BP neural network algorithm proposed in this paper can meet the recognition needs.3.A fusion algorithm based on cloud model based on step-by-step D-S evidence theory and weighted average.The traditional D-S evidence theory fusion algorithm is difficult to assign initial probability.To solve this problem,a cloud model is introduced to complete the initial probability assignment.At the same time,aiming at the problem of poor fusion of high-conflict evidence in traditional DS evidence theory,the advantages and disadvantages of several improved algorithms and the identification requirements of this paper are compared.A method combining the stepwise DS evidence theory fusion algorithm and weighted average algorithm is proposed.The problem was resolved and the required identification requirements were met.4.Evaluation experiment of vehicle bottom dangerous goods identification system.In order to verify the feasibility and different characteristics of the two fusion algorithms,several experimental sites and platforms for different positions of the bottom of the vehicle and different dangerous goods were set up.The two fusion algorithms were used to detect dangerous goods.The results show that the two multi-channel information fusion algorithms proposed in this paper can meet the requirements of vehicle dangerous goods identification,and the danger classification is accurate.The difference is that when the multi-sensor information conflict is small,the fusion algorithm based on the cloud model’s step-by-step D-S evidence theory and weighted average is faster;when the conflict is larger,the improved BP neural network fusion algorithm is faster. |
