This paper proposes an incremental object classification based on parts detected in a sequence of noisy range images. Primitive parts are jointly tracked and detected as probabilistic bounding-boxes using a particle filter which accumulates the information of the local structure over time. A voting scheme is presented as a procedure to verify structure of the object, i.e. the desired geometrical relations between the parts. This verification is necessary to disambiguate object parts from potential irrelevant parts which are structurally similar. The experimental results obtained using a mobile robot in a real indoor environment show that the presented approach is able to successfully detect chairs in the range images.