Sistema ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

 

Tesi etd-03142018-174808


Tipo di tesi
Tesi di laurea magistrale
Autore
CIAMPI, LUCA
URN
etd-03142018-174808
Titolo
Designing and Testing a Deep Learning Based System for Counting Vehicles with Cameras
Struttura
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
COMPUTER ENGINEERING
Commissione
relatore Gennaro, Claudio
Parole chiave
  • Convolutional Neural Networks
  • Deep Learning
  • Counting
  • CNN
Data inizio appello
07/05/2018;
Consultabilità
parziale
Data di rilascio
07/05/2021
Riassunto analitico
The counting problem is the estimation of the number of objects instances in still images or video frames. It involves various domains, for instance, counting cells or bacteria from microscopic images, estimate the number of people in a crowd scene or counting animals in ecological surveys with the intention of monitoring the population of a certain region. In this thesis work, we deal with the specific problem of counting vehicles in a car park, using frames captured by various different cameras.

In principle, the key idea behind objects counting is very simple: density times area. However, objects are not regular across the scene. They cluster in certain regions and are spread out in others. Another factor of complexity is represented by perspective distortions created by different camera viewpoints in various scenes, resulting in large variability of scales of objects. Others challenges points to be considered are inter-object occlusions, high similarity of appearance between objects and background elements, different illumination, and low image quality.

Objects counting has been tackled as a computer vision problem using many techniques: almost all state-of-the-art approaches make an extensive use of Convolutional Neural Networks (CNNs). In this work, as a first step, we have examined and tested some of these existing CNNs: Multi-column CNN, Hydra CNN, CrowdNet, YOLO9000, mAlexNet and Mask R-CNN}.
Tests are performed using Counting CNRPark-EXT Dataset, a new dataset specifically created in this thesis work for the counting task, that extends the existing CNRPark-EXT Dataset, a collection of images from the parking lots in the campus of the National Research Council (CNR) in Pisa. Then, as a second step, a fine tuning of the best solutions previously identified is performed.

We finally show how the results obtained using these fine-tuned networks outperform all the others and how a real car counting system could be implemented, evaluating the number of vehicles and the car park occupancy status from a camera video stream.
File