Roberto Ugolotti

Particle Swarm Optimization and Differential Evolution for model-based object detection

Automatically detecting objects in images or video sequences is one of the most relevant and frequently tackled tasks in computer vision and pattern recognition. The starting point for this work is a very general model-based approach to object detection. The problem is turned into a global continuous optimization one: given a parametric model of the object to be detected within an image, a function is maximized, which represents the similarity between the model and a region of the image under investigation. In particular, in this work, the optimization problem is tackled using Particle Swarm Optimization (PSO) and Differential Evolution (DE). We compare the performances of these optimization techniques on two real-world paradigmatic problems, onto which many other real-world object detection problems can be mapped: hippocampus localization in histological images and human body pose estimation in video sequences. In the former, a 2D deformable model of a section of the hippocampus is fit to the corresponding region of a histological image, to accurately localize such a structure and analyze gene expression in specific sub-regions. In the latter, an articulated 3D model of a human body is matched against a set of images of a human performing some action, taken from different perspectives, to estimate the subjects posture in space. Given the significant computational burden imposed by this approach, we implemented PSO and DE as parallel algorithms within the nVIDIA(TM) CUDA computing architecture.

Automatic hippocampus localization in histological images using Differential Evolution-based deformable models

In this paper, the localization of structures in biomedical images is considered as a multimodal global continuous optimization problem and solved by means of soft computing techniques. We have developed an automatic method aimed at localizing the hippocampus in histological images, after discoveries indicating the relevance of structural changes of this region as early biomarkers for Alzheimers disease and epilepsy. The localization is achieved by searching the parameters of an empirically-derived deformable model of the hippocampus which maximize its overlap with the corresponding anatomical structure in histological brain images. The comparison between six real-parameter optimization techniques (Levenberg-Marquardt, Differential Evolution, Simulated Annealing, Genetic Algorithms, Particle Swarm Optimization and Scatter Search) shows that Differential Evolution significantly outperforms the other techniques in this task, providing successful localizations in 90.9% and 93.0% of two test sets of real and synthetic images, respectively.

Multi-sensor system for detection and classification of human activities

This paper describes a novel system for detecting and classifying human activities based on a multi-sensor approach. The aim of this research is to create a loosely structured environment, where activity is constantly monitored and automatically classified, transparently to the subjects who are observed. The system uses four calibrated cameras installed in the room which is being monitored and a body-mounted wireless accelerometer on each person, exploiting the features of different sensors to maximize recognition accuracy, improve scalability and reliability. The algorithms on which the system is based, as well as its structure, are aimed at analyzing and classifying complex movements (like walking, sitting, jumping, running, falling, etc.) of potentially multiple people at the same time. Here, we describe a preliminary application, in which action classification is mostly aimed at detecting falls. Several instances of a hybrid classifier based on Support Vector Machines and Hierarchical Temporal Memories, a recent bio-inspired computational paradigm, are used to detect potentially dangerous activities of each person in the environment. If such an activity is detected and if the person “in danger” is wearing the accelerometer, the system localizes and activates it to receive data and then performs a more reliable fall detection using a specifically trained classifier. The opportunity to turn on the accelerometer on-demand makes it possible to extend its battery life. Besides and beyond surveillance, this system could also be used for the assessment of the degree of independence of elderly people or, in rehabilitation, to assist patients during recovery.

libCudaOptimize: an open source library of GPU-based metaheuristics

Evolutionary Computation techniques and other metaheuristics have been increasingly used in the last years for solving many real-world tasks that can be formulated as optimization problems. Among their numerous strengths, a major one is their natural predisposition to parallelization. In this paper, we introduce libCudaOptimize, an open source library which implements some metaheuristics for continuous optimization: presently Particle Swarm Optimization, Differential Evolution, Scatter Search, and Solis&Wets local search. This library allows users either to apply these metaheuristics directly to their own fitness function or to extend it by implementing their own parallel optimization techniques. The library is written in CUDA-C to make extensive use of parallelization, as allowed by Graphics Processing Units. After describing the library, we consider two practical case studies: the optimization of a fitness function for the automatic localization of anatomical brain structures in histological images, and the parallel implementation of Simulated Annealing as a new module, which extends the library while keeping code compatibility with it, so that the new method can be readily available for future use within the library as an alternative optimization technique.

Real-Time GPU based road sign detection and classification

This paper presents a system for detecting and classifying road signs from video sequences in real time. A model-based approach is used in which a prototype of the sign to be detected is transformed and matched to the image using evolutionary techniques. Then, the sign detected in the previous phase is classified by a neural network. Our system makes extensive use of the parallel computing capabilities offered by modern graphics cards and the CUDA architecture for both detection and classification. We compare detection results achieved by GPU-based parallel versions of Differential Evolution and Particle Swarm Optimization, and classification results obtained by Learning Vector Quantization and Multi-layer Perceptron. The method was tested over two real sequences taken from a camera mounted on-board a car and was able to correctly detect and classify around 70% of the signs at 17.5 fps, a similar result in shorter time, compared to the best results obtained on the same sequences so far.

A comparative study of three GPU-based metaheuristics

In this paper we compare GPU-based implementations of three metaheuristics: Particle Swarm Optimization, Differential Evolution, and Scatter Search. A GPU-based implementation, obviously, does not change the general properties of the algorithms. As well, we give for granted that GPU-based implementation of both algorithm and fitness function produces a significant speed-up with respect to a sequential implementation. Accordingly, the main goal of this work has been to fairly assess the efficiency of the GPU-based implementations of the three metaheuristics, based on the statistical analysis of the results they obtain in optimizing a benchmark of twenty functions within a prefixed limited time.

Automatic segmentation of hippocampus in histological images of mouse brains using deformable models and random forest

We perform a two-step segmentation of the hippocampus in histological images. First, we maximize the overlap of an empirically-derived parametric Deformable Model with two crucial landmark sub-structures in the brain image using Differential Evolution. Then, the points located in the previous step determine the region where a thresholding technique based on Otsus method is to be applied. Finally, the segmentation is expanded employing Random Forest in the regions not covered by the model. Our approach showed an average segmentation accuracy of the 92.25% and 92.11% on test sets comprising 15 real and 15 synthetic images, respectively.

Visual Search of Neuropil-Enriched RNAs from Brain In Situ Hybridization Data through the Image Analysis Pipeline Hippo-ATESC

Motivation RNA molecules specifically enriched in the neuropil of neuronal cells and in particular in dendritic spines are of great interest for neurobiology in virtue of their involvement in synaptic structure and plasticity. The systematic recognition of such molecules is therefore a very important task. High resolution images of RNA in situ hybridization experiments contained in the Allen Brain Atlas (ABA) represent a very rich resource to identify them and have been so far exploited for this task through human-expert analysis. However, software tools that may automatically address the same objective are not very well developed. Results In this study we describe an automatic method for exploring in situ hybridization data and discover neuropil-enriched RNAs in the mouse hippocampus. We called it Hippo-ATESC (Automatic Texture Extraction from the Hippocampal region using Soft Computing). Bioinformatic validation showed that the Hippo-ATESC is very efficient in the recognition of RNAs which are manually identified by expert curators as neuropil-enriched on the same image series. Moreover, we show that our method can also highlight genes revealed by microdissection-based methods but missed by human visual inspection. We experimentally validated our approach by identifying a non-coding transcript enriched in mouse synaptosomes. The code is freely available on the web at http://ibislab.ce.unipr.it/software/hippo/.

Differential evolution based human body pose estimation from point clouds

This paper describes a method to estimate the body pose of a human from the point cloud obtained from a depth sensor. It uses Differential Evolution to find the best match between a candidate pose, represented by an instance of a 42-parameter articulated model of a human, and the point cloud. The results, compared to other four state-of-the art methods on a publicly available dataset, show that the method has good ability to estimate the pose of a person and to track him in video sequences. The entire method, from Differential Evolution to fitness computation, is run on nVIDIA GPUs. Thanks to its massively parallel implementation in CUDA-C, it produces pose estimates in real time.

GPU-Based Point Cloud Recognition Using Evolutionary Algorithms

In this paper, we describe a method for recognizing objects in the form of point clouds acquired with a laser scanner. This method is fully implemented on GPU and uses bio-inspired metaheuristics, namely PSO or DE, to evolve the rigid transformation that best aligns some references extracted from a dataset to the target point cloud. We compare the performance of our method with an established method based on Fast Point Feature Histograms (FPFH). The results prove that FPFH is more reliable under simple and controlled situations, but PSO and DE are more robust with respect to common problems as noise or occlusions.