The traditional Convolutional Neural Network (CNN) cannot directly process point cloud data, and the point cloud data must be converted into a multi-view or voxelized grid, which leads to a complicated process and low point cloud recognition accuracy. Aiming at the problem, a new point cloud classification and segmentation network called Linked-Spider CNN was proposed. Firstly, the deep features of point cloud were extracted by adding more Spider convolution layers based on Spider CNN. Secondly, by introducing the idea of residual network, short links were added to every Spider convolution layer to form residual blocks. Thirdly, the output features of each layer of residual blocks were spliced and fused to form the point cloud features. Finally, the point cloud features were classified by three-layer fully connected layers or segmented by multiple convolution layers. The proposed network was compared with other networks such as PointNet, PointNet++ and Spider CNN on ModelNet40 and ShapeNet Parts datasets. The experimental results show that the proposed network can improve the classification accuracy and segmentation effect of point clouds, and it has faster convergence speed and stronger robustness.
Focusing on the issue that the Signal-to-Clutter-and-Noise Ratio (SCNR) of echo signal is low when cognitive radar detects extended target, a waveform design method based on SCNR was proposed. Firstly, the relation between the SCNR of cognitive radar echo signal and the Energy Spectral Density (ESD) of transmitted signal, was gotten by establishing extended target detection model other than previous point target model; secondly, according to the maximum SCNR criterion, the global optimal solution of the transmitted signal ESD was deduced; finally, in order to get a meaningful time-domain signal, ESD was synthesized to be a constant amplitude based on phase-modulation after combining with the Minimum Mean-Square Error (MMSE) and iterative algorithm, which met the emission requirements of radar. In the simulation, the amplitude of time-domain synthesized signal is uniform, and its SCNR at the output of the matched filter is 19.133 dB, only 0.005 dB less than the ideal value. The results show that not only can the time-domain waveform meet the requirement of constant amplitude, but also the SCNR obtained at receiver output can achieve the best approximation to the ideal value, and it improves the performance of the extended target detection.
Aiming at the defects of resolving large scale 0-1 knapsack problem with existed algorithm, an Improved Artificial Bee Colony algorithm based on P Systems (IABCPS) was introduced in this paper. The idea of Membrane Computing (MC), polar coordinate coding and One Level Membrane Structure (OLMS) was used by IABCPS. Evolutionary rules of improved Artificial Bee Colony (ABC) algorithm and transformation or communication rules in P systems were adopted to design its algorithm. The limit of number of trials "limit" was adjusted to keep the balance of exploitation and exploration. The experimental results show that IABCPS can find the optimum solutions in solving small scale knapsack problems. In solving a knapsack problem with 200 items, compared with Clonal Selection Immune Genetic Algorithm (CSIGA), IABCPS increases the average results by 0.15% and decreases variance by 97.53%; compared with ABC algorithm, IABCPS increases the average results by 4.15% and decreases variance by 99.69%. The results demonstrate that IABCPS has good ability of optimization and stability. Compared with Artificial Bee Colony algorithm based on P Systems (ABCPS) in solving large scale knapsack problem with 300, 500, 700 and 1000 items respectively, IABCPS increases the average results by 1.25%,3.93%,6.75% and 11.21%, and the ratio of the variance and the number of experiments keeps in single digits. It shows its strong robustness.
The performance of the Graph-based Semi-Supervised Learning (GSSL) method based on one graph mainly depends on a well-structured single graph and most algorithms based on multiple graphs are difficult to be applied while the data has only single view. Aiming at the issue, a Graph Transduction via Alternating Minimization method based on Multi-Graph (MG-GTAM) was proposed. Firstly, using different graph construction parameters, multiple graphs were constructed from data with one single view to represent data point relation. Secondly,the most confident unlabeled examples were chosen for pseudo label assignment through the integration of a plurality of map information and imposed higher weights to the most relevant graphs based on alternating optimization,which optimized agreement and smoothness of prediction function over multiple graphs. Finally, more accurate labels were given over the entire unlabeled examples by combining the predictions of all individual graphs. Compared with the classical algorithms of Local and Global Consistency (LGC), Gaussian Fields and Harmonic Functions (GFHF), Graph Transduction via Alternation Minimization (GTAM), Combined Graph Laplacian (CGL), the classification error rates of MG-GTAM decrease on data sets of COIL20 and NEC Animal. The experimental results show that the proposed method can efficiently represent data point relation with multiple graphs, and has lower classification error rate.
Concerning the problem that the weak target might be covered by the range side-lobes of the strong one and the range side-lobes could only be suppressed to a certain value, an improved Kalman-Minimum Mean-Square Error (K-MMSE) algorithm was proposed in this paper. This algorithm combined the Kalman filter with the Minimum Mean-Square Error (MMSE), and it was an effective method for suppressing range side-lobes of adaptive pulse compression. In the simulation, the proposed algorithm was compared with the traditional matched filter and other improved matched filters such as MMSE in a single target or multiple targets environments, and then found that the side-lobe levels, the Peak-SideLobe Ratio (PSLR) and Integrated SideLobe Ratio (ISLR) of the Point Spread Function (PSF) were all decreased obviously in comparison with the previous two methods. The simulation results show that the method can suppress range side-lobes well and detect the weak targets well either under both the condition of a single target and the condition of multiple targets.
A classification method based on trinocular stereovision, which consisted of geometrical classifier and color classifier, was proposed to autonomously guide vehicles on unstructured terrain. In this method, rich 3D data which were taken by stereovision system included range and color information of the surrounding environment. Then the geometrical classifier was used to detect the broad class of ground according to the collected data, and the color classifier was adopted to label ground subclasses with different colors. During the classifying stage, the new classification data needed to be updated continuously to make the vehicle adapt to variable surrounding environment. Two broad categories of terrain what vehicles can drive and can not drive were marked with different colors by using the classification method. The experimental results show that the classification method can make an accurate classification of the terrain taken by trinocular stereovision system.