Monocular Vision Based Environment Cognition And Mapping For Robots

With the development of intelligent technology, more and more intelligent service robots will be applied to people’s daily life. To serve people, a robot must have the competence of environment cognition, such as scene understanding and object recognition. Vision is the most important way for people to perceive environment. It is well known that more than 80 percent information is acquired by vision. Camera is the main sensor of robot systems thanks to its small size and relatively low price. In recent years, many impressive visual systems have been proposed, however, most of them are based on low level features, which leads to lack semantic meaning. Consequently, robots can not perform higher level intelligent tasks. In this thesis, in order to improve cognitive ability of robots, we have studied on scene classification, fine-grained recognition, object tracking, and mapping. The main contents are as follows:(1) A scene classification method based on mid-level features represented by sparse coding is proposed. First, since the information of an image is mainly concentrated in the contour regions. An adaptive threshold Canny edge detection algorithm is applied to extract the contours of the image, and the candidate mid-level image blocks are obtained using a sliding window strategy. Then, a feature representation method based on sparse coding is adapted to describe the mid-level image blocks. Finally, the most representative blocks of each scene are discovered by using the K-means clustering method with the critera of purity and identification. And the SVM classifier is used to perform classification. The experimental results on different datasets show that the proposed method can obtain higher recognition rate than other counterparts.(2) A novel Weighted Coefficient Deformable Parts Model(WCDPM) is presented for fine-gained recognition. First, the Latent SVM is utilized to train the Deformable Parts Model(DPM). Then, a learning method of weighted coefficient for each component of the DPM is proposed. Different to the traditional DPM, the WCDPM can highlight the contributions of different parts in the same kind of objects, which is helpful to fine-grained recognition. Experimental results demonstrate that the proposed method performs better than other counterparts in terms of recognition rate.(3) In order to track moving objects, a temporal-spatial context tracking method based on normalized color features is proposed. First, the normalized color histogram features of the object are calculated. Then, the temporal and spatial relation between the object and the local context is established based on the Bayesian framework. Experimental results show that the proposed method obtain better performance than its counterparts without obvious increase of computational burden.(4) A hierarchical map of the environment is built for indoor localization and navigation. The hierarchical map consists of two layers: topological layer and semantic layer. The topological layer provides information for self-localization and path planning using image retrieval algorithm and pose estimation algorithm respectively; while the semantic layer provides scene and object information for object localization using scene and object classification algorithms. A tracking approach is also used to track the moving objects when mapping. A demo system is implemented based on the P3-DX platform to evaluate the proposed approaches.