In order to improve the automatic detection accuracy of urban buildings, a new automatic detection technique of urban buildings is proposed with combination of laser radar scanning and optical imaging technique. The proposed technique delivers the point cloud generated by laser radar scanning to the auto-encoder to down sampling and extract features, and the optical image pixels are delivered to auto-encoder to extract the spatial features, different feature sets are fused based on the graph model of features. Then, a semi-supervised classifier is designed based on convolutional neural network, the classifier is used to recognize different regions of buildings, ground and vegetation. Experimental results show that the proposed technique can detect the regions of urban buildings accurately, it also provides fundamental basis for smart cities development.