Research On Obstacle Recognition Based On Vision For Deicing Robot On Voltage Transmission Line

Obstacle recognition is one of the key techniques of autonomous deicing robot on high voltage transmission line. The obstacles such as counterweight, strain clamp and suspension on high voltage transmission line should be effectively recognized for deicing robot to autonomously cross obstacles. According to the structure of 220kV transmission line, a series of methods for obstacle recognition based on vision are put forward in this thesis. The proposed methods don't require structure constraint and can achieve good obstacle classification results. The main studies in this paper are as follows.1) The obstacle images are pretreated firstly, and then the edges of obstacle images are detected by using wavelet modulus maximum algorithm and canny algorithm respectively. The result shows that the edges of obstacle images which detected by wavelet modulus maximum algorithm which have a stronger anti-jamming capability have a better edge detection performance than canny algorithm.2) The moment features of obstacle's edge images are selected as obstacle classifier input vector. Then the united moments and the wavelet moments are calculated and the both moment features are reduced dimension and optimized by using sub-optimal search algorithm.3) Obstacle classification methods based on neural network are researched. First of all, a multilayer feed-forward neural network and a wavelet neural network both based on BP algorithm are proposed. Since BP algorithm has slow convergence and is easy tendency to partial optimization, the particle swarm optimization (PSO) algorithm with a faster convergence and a stronger global search capability is introduced to replace the BP algorithm to training the wavelet neural network, then a wavelet network based on PSO is established. However, PSO algorithm may still fall into partial optimization, drawing on the idea of sudden jump in simulated annealing, at last, a improved PSO wavelet network based on simulated annealing algorithm is developed and demonstrates excellent classification performance.4) Taking into account the support vector machine (SVM) is more suitable for small sample pattern recognition problem, obstacle classification methods based on SVM are also studied. To begin with, a SVM classifier based on grid search and cross-validation for parameter optimization is proposed. Secondly, taking the blindness of the grid search into account, a SVM classifier based on PSO and cross-validation for parameter optimization is proposed and achieves excellent classification performance. Additionally, in order to eliminate the relevance of the data, the momont features inputed to both SVM classifiers are processed through independent component analysis using FastICA algorithm. This improved the classification performance of both classifiers greatly.