Three lower level layers of the hierarchical machine perception system are described, and experimental results are provided. The system is based on agile camera and aimed at automatic recognition of simple geometric scenes, made of crafted manipulation objects. The primary process of grouping the edgelets on the image is described. Two methods of edgelet grouping are compared, namely a Hough transform based method and a method based on LSR. Some modifications to both methods have been proposed. Next, a process of perceptual grouping of elementary features, short segments, is described. Three approaches to feature fusion are addressed: one based on crisp, geometrical, heuristic conditions, the second based on fuzzified conditions and the third one based on Dempster-Shafer theory. The results obtained with the three methods are compared and presented on the example image of a real scene. Higher layer issues, namely a problem of contour closing and of separating particular objects (in the case of occlusions) are briefly discussed. Some preliminary results in that area are reported.