Conception et mise en place d'une solution Deep Learning pour la détection des véhicules dans les

Abstract #

Data fusion has been widely studied in the field of artificial intelligence. Information

is generally considered imperfect. Therefore, the combination of several heterogeneous

information sources can lead to a more global, more complete and more accurate information

[1].

In this regard, the Société Nationale des Autoroutes du Maroc (ADM) is aware of the

importance and purposes of this merger to help improve the performance of systems for

predicting and monitoring the flow of highway traffic, and has decided to exploit all of its

data sources from various data providers. These providers include surveillance cameras and

transactional data.

In this context, this memoir is included, which aims to focus firstly on the use of video data

for the detection and counting of vehicles in real time and secondly to merge these data with

the transactional data to help improve the prediction of highway traffic flow.

To achieve this goal, convolutional neural networks are being studied with the aim of

developing a detection model, capable of recognizing vehicles in video streams with high

precision and fast computing time. YOLOV4 is the vehicle detection model that has been

studied and optimized throughout this work. The DeepSORT tracking algorithm has been

studied and associated with the YOLOV4 detection model in a final model that will be

responsible for the detection and counting of vehicles in particular areas, through different

videos in real time.

In the second part of the project, we are interested in, the fusion of transactional data and

video data as a support to improve the prediction of the flow of motorway traffic.

For this, we have studied the theory of evidence for modeling, combination and decision. This

theory makes it possible to represent both the uncertainty and the imprecision of the

information obtained from each source, by assigning weights for each data source which

apply not only to a single condition but to a set of conditions.

Keywords: Deep Learning, convolutional neural network (CNN), real-time object detection,

real-time object tracking, multiple object tracking, data fusion.