A Multi-view skeletal tracking system

This projct aims at developing a system for pose estimation based on multiple observations of the same person from different viewpoints. We exploit a pose estimation agorithm for analyzing the images provided by the different cameras separately. The final outcome is obtained by triangulating all the views in the 3D space, and merging all the triangulated points by means of a geometrical method.
Triangulated body pose

The Themobot project: Autonomous robotic system for thermographic detection of cracks

Crack detection in parts of complex geometry is often done manually using magnetic particle inspection. The Thermobot project combines robotics and thermography to replace this decades-old method. Using the inspection of a whole crankshaft as an example a prototype system will be developed that can inspect complex parts, mainly targeting at the automotive industry but also at aerospace applications.
I was involved in the research and development of algorithms for automatic detection of cracks in metal parts and material intrusions in carbon fiber parts. Such algorithms are able to deal with the different orientations under which a part can be seen, and distinguish between artifacts and real defects.
In the following image you can see the how the heat diffusion is modified by the presence of a crack. The system is able to detect this and highlight the anomalous heat diffusion (indicated by the red dots)
Heat diffusion measurement
More information can be found in the Thermobot web page.

Video surveillance of big events

This project, in cooperation with the University of Lincoln (UK), aims at creating smart systems for automating video surveillance of large groups of people attending big events, like football matches, concerts, and car races. We are currently addressing novel ways for detecting crowds of people, a task that cannot be solved exploiting common people detection systems.

Here you can find a typical scenario:
Crowd in a venue
When a reliable crowd detection and localization system will be available, we will face more advanced tasks, like people tracking in crowded environment, and cooperative vision, in order to exploit the large number of cameras usually installed in a venue.

Cooperative vision and multimedia middlewares

All the software developed for video surveillance applications is included in a framework: this way all low-level details (like video acquisition) can be neglected, and communication among software modules is standardized.

The middleware we are currently using is NMM, developed by Motama. Some modules I worked on are:
  • Fall detection system (in cooperation with Alberto Salamone)
  • Motion blob tracking for video surveillance applications
  • Omnidirectional image unwarping
  • PTZ camera control modules
The fall detection system is aimed at helping elderly people that like to live alone: the automatic system can guarantee an optimal privacy level, together with a constant monitoring, without the need for wearing specific equipment for issuing alerts. A prototype of this system is currently installed in Castelnuovo del Garda, developed in the project SafeHome (sponsored by Veneto region).
Fall detector

The Omnidome integrated video sensor

Omnidome is a vision sensor composed by an omnidirectional camera and a Pan-Tilt-Zoom (PTZ) unit:
I exploited this sensor for developing a system able to detect people moving around and recording a shot of each one. Here you can see a portion of the unwarped omnidirectional image, and a face shot:
Panoramic image Face shot

Past projects

DARPA Urban Challenge

The DARPA Urban Challenge is a race for autonomous vehicles, that took place in November 2007 in Victoriville, California. I took part at the event as a member of Team Terramax: I developed a system for detecting ovetaking vehicles (a sort of automatic rearview mirror). This is an example:
Rear View TerraMax
Here you can see our vehicle:

Detection of dangerous situations for pedestrians in urban environment

This project presents a novel approach to the task of pedestrian detection. The basic idea is to perform a preliminary analysis of the laserscanner data in order to detect possible dangerous situations, and drive the pedestrian detection system on them. I have been working on the system analyzing laserscanner data, that provides a map of the area in front of the vehicle, as well as a list of areas that should be checked by the pedestrian detector.

Pedestrian contour extraction in a tetravision system

Pedestrian detection is one of the most challenging tasks in the field of computer vision, due to the high variety of poses and colors in which people can appear in the images. A reliable detection systems should therefore be based on the analysis of a high number of featurs, like shape. In this project an acquisition system called tetravision has been used, that is composed of two stereo camera pairs, one in the visible range, and the other in the far infrared domain. To improve results, I worked on the shape extraction of obstacles detected by stereo processing in the infrared images: extracted shapes are used to distinguish between pedestrians and general obstacles. The technique of active contours (snake) has been used: in the following, you can find a sample image, in which contour extraction of a generic obstacle and a pedestrian are shown:
Contorno pedoni

Distant pedestrian detection system

Pedestrian detection systems are usually focused on the urban environment, where the majority of the accidents involving pedestrians happens. However, few pedestrians can also be found outside cities, and these situations turn out to be particularly dangerous, due to the lack of artificial lighting during the night. On the other hand, scenes acquired outside the urban environment are easier to analyze, and it is then possible to achieve very high performance. In this project, I developed technique meant for detecting very distant pedestrians in infrared images, in which pedestrians can be seen without the need for lighting. Due to the low resolution of this kind of images, the technique is based on probabilistic models; in the following you can find some examples of detection:
Pedoni lontani rilevati Pedoni lontani rilevati

Automatic calibration of a stereo camera pair by using zebra crossing signs

This projects aims at detecting pedestrians crossing a street by using a stereo camera pair. Since both cameras observe a known pattern (the zebra crossing signs), it has been possible to exploit it in order to compensate both lens distortion and perspective effect. Here you can see two sample images: in the first one, just the original image is shown, while in the second one it is possible to see the same image after distortion compensation and perspective removal:
Strisce pedonali rilevate Immagine dedistorta

Design downloaded from free website templates.