Research Of Robot's Environmental Perception Based On Multi-sensor Fusion

Today, the research and application areas on Intelligent Mobile Robot gradually expand, and the robot's operating environment has become increasingly complex and diverse. In order to obtain better learning ability, a modern intelligent robot's context-aware system, relying on a single sensor's information, is unable to meet the needs of performing complex tasks. The context-aware system should use multi-sensors to obtain rich observation data, and do real time process, optimization and integration, to achieve state estimation, environmental detection, target identification and tracking, behavioral intentions analysis, situation assessment and other functions. To this end, since the1970s, multi-sensor information fusion technology has been rising rapidly, and then becoming a high-level key technology concerned by multi-disciplinary and many areas.Data association of multi-sensors'information with complex non-linear relationship is one of the most ticklish issues in robot's context-aware system. This thesis introduces the kernel methods into solving this problem. This dissertation will discuss the problem of how to fuse multi-sensor information by kernel methods, and thus to realize robot environment perception. The main aspects of work include:1, Propose a new idea that using kernel methods for pattern recognition to fuse multi-sensor information. Lucubrate two modes of how to apply kernel methods to multi-sensor information fusion, one is the application model in multi-sensor information's feature extraction link and the other is the application mode in SVM multi-class discriminator's design link. To counter two key issues in the application of kernel methods--the selection of kernel function and its parameters, this paper analyzes and compares the performance of several common kernel functions from a theoretical perspective, and gives an approach to select the optimal kernel parameter and penalty factor.2, According to the theory of kernel methods and standard SVM classifier theory, design a multi-class discriminator based on kernel methods, realize the multi-sensor information's feature fusion in the discriminator. Conduct real world experiment relying on the AS-RF Robot as the experimental platform, and complete the pose prediction and discrimination of the robot. The experimental results show that, in a simple structure and ideal environment, robot can achieve90%or higher discriminant precision.3, Combined kernel methods and typical feature extraction algorithms to extract the main environment features from multi-sensor data using kernel principal component analysis (KPCA) feature extraction method. Then, based on the extracted features of the environment, and combined with particle filtering algorithm, the robot's self-localization experiment is conducted, and gets an acceptable result. The experimental results indicate that, when the particle number is large (such as2000or2500particles), it can achieve about80%success rate of positioning. By contrast with several similar experiments, we can infer that this experiment does provide good practical reference value.