M. Accame

Using machine learning techniques in real-world mobile robots

Applying machine learning techniques can help mobile robots meet the need for increased safety and adaptivity that real world operation demands. The techniques also facilitate robot to user communication. Using these techniques, we built increasingly abstract representations of a robots perceptions and actions. This produced a symbolic description of what the robot knows and can do. Because this task is fairly complex, we first identified those subproblems that a learning method can solve efficiently, and isolated those with good classical solutions. Also, for a robot to solve a complex problem, we had to find solutions for several learning tasks. We identified these learning tasks and the learning techniques appropriate for their solution. To evaluate our approach, we used the mobile robots Priamos and Teseo. >

Hierarchical block matching for disparity estimation in stereo sequences

The paper focuses on a new method for block based disparity estimation (BBDE) that is specially suited for real time coding of stereo sequences. The estimation is performed by exploiting a preliminary disparity field, obtained from the coarsest level of a multiresolution pyramid, and its successive refinements in the finer ones. A strong correction strategy for wrong estimates has been implemented by using an appropriate propagation of disparity vectors from one level to the next one and an adaptive search range. Finally, a variable-resolution disparity field is achieved stopping the propagation at an intermediate level of resolution, if a vector allows a good reconstruction quality of its block.

Hierarchical progressive image coding controlled by a region based approach

This paper presents a progressive image coding technique in which a rough preview image is enhanced with the zerotree encoding of the residual image in the wavelet domain. The preview image is achieved with a region based approach that gives a basic description of the scene with a very high compression ratio. As a second stage, the residual between the original image and the preview is decomposed using a wavelet transform, and it is encoded for a progressive transmission. The wavelet coefficients are selected from the coarsest to the finest level of resolution, using a mask obtained from the image preview. Such a mask sets the importance of the coefficients on the basis of their distance from the contours of the regions of the preview. A modified zerotree approach is used to progressively encode the weighted coefficients until significant texture is introduced in the image and the required quality is achieved. Moreover, if an image with very good quality is required, the resulting reconstruction error is encoded by means of vector quantization and transmitted over the channel. The proposed method is compared to progressive JPEG, and it gives similar results at low compression ratios and far better results at higher ones.

Hierarchical motion estimator (HME) for block-based video coders

A hierarchical block-based motion estimation strategy for video sequence coding is proposed, which allows to reduce both the computational load and the number of transmitted motion vectors. The estimation is performed by first computing a coarse motion field at the lowest resolution of a Gaussian pyramid and then propagating and refining it till the maximum resolution. the main innovations concern the adaptive vector propagation and updating strategy, which allow an efficient recovery from wrong estimates while maintaining a low computational load. The advantages of the proposed method are manifold: first, it is able to decrease the computation of even two orders magnitude for an equivalent compensation; second, it generates a lower number of motion vectors, resulting in a lower information rate; third, it produces a more homogeneous motion field, thus easing a successive segmentation.

Efficient labeling procedures for image partition encoding

Abstract A new approach to the lossless encoding of an image partition is presented. The segmented image is first described by a quadtree, whose leaves are then grouped by a labeling procedure to represent any configuration. Label assignment exploits the ‘four colors’ theorem, thus allowing to encode each label with 2 bits. A simple and efficient label assignment algorithm is also proposed, which further reduces the code entropy. The proposed representation is an effective alternative to region- and edge-based partition encoders, used in II generation image coders.

An integrated approach to block based motion estimation for video coding

In video communication systems motion compensation techniques allow one to remove temporal redundancy among frames in order to achieve reductions in the bit rate. Block based approaches are widely used for such a task, since they offer a good balance between performance and complexity; however, the resulting motion fields still present spatial and temporal redundancies, the exploitation of which would allow a greater bit rate reduction. The motion estimation method proposed attains such a goal, maintaining the block based partitioning of the resulting field. It uses a spatio-temporal prediction followed by a lossy coding of the residual field by vector quantization. The prediction is based on autocompensation, enhanced with the aid of some additional parameters so that it provides a stationary error field. The statistical distribution of such a residual field allows vector quantization to perform efficiently, even when the method is applied to video sequences with extremely varying motion activities. Both the prediction and the vector quantization simply use an index for each block to code its motion vector. Finally, the indexes are spatially organized in two single images that are coded with a region based technique.

Adaptive 3D interpolation and two-source video coding

A new strategy for video coding at very low bit rates is proposed in the paper, where an image sequence is encoded in a two-source way. The main difference from state-of-the-art approaches lies in the use of a 3D spatio-temporal interpolation, based on a non-uniform grid (frames are processed in overlapped blocks), in order to allocate the information in an adaptive way. Experimental results show a good fidelity of the decoded sequences, without blocking effect.<<ETX>>

ANN-driven edge point selection criterion

This paper presents a new strategy that exploits artificial neural networks (ANNs) for a direct selection of edge points from an image. First, a spatial filtering for edge enhancement (the Canny filter) is used to obtain a set of candidate edge points which turn out to be the local maxima of the filtered image (MPS). A preliminary coarse selection of these points that exploits neighborhood information is performed to produce an extended pseudo-edges set (PES). Then, a features vector is extracted from the PES and is used by a neural classifier to decide whether or not a point belongs to the target edge set (TES).

Disparity Estimation for Stereo Sequences Based on Adaptive Size Hierarchical Block Matching

The paper focuses on a new method for block based disparity estimation (BBDE) that is specially suited for real time coding of stereo sequences. The estimation is performed by exploiting a preliminary disparity field, obtained from the coarsest level of a multiresolution pyramid, and its successive refinements in the finer ones. A strong correction strategy for wrong estimates has been implemented by using an appropriate propagation of disparity vectors from one level to the next one and an adaptive search range. Finally, a variable-resolution disparity field is achieved stopping the propagation at an intermediate level of resolution, if a vector allows a good reconstruction quality of its block.

Learning to Control a Visual Sensing System

This chapter shows how to learn the extraction of visual primitives from an image. The considered primitives are edges that provide visual information about the environment in which the robot work. Their extraction can be obtained by means of four modules working in chain. The result is that raw data acquired with a camera is transformed into semantically relevant data that take information about the robot environment. In this context, Machine Learning is used to tune image acquisition and edge extraction, so that the Visual Sensing System (VSS) adapts itself to a dynamic environment. An experiment is presented that shows how the VSS is able to find the exact location of a door acquired by the robot camera.