الملخص الإنجليزي
ABSTRACT :
In response to the pressing need for effective crowd management solutions for robotics,
this study delves into the creation of a deep learning-based scene and face recognition
system tailored for robotic crowd management. The primary objective is to develop a
system that not only exhibits high levels of accuracy and efficiency but also adheres to
ethical principles. This is achieved through the integration of deep learning models with
multi-modal fusion approaches, allowing for the swift and comprehensive assessment
of optical, auditory, and depth data to identify potential hazards and maintain crowd
control. At the heart of this research lies the meticulous gathering and preparation of
data. Establishing a broad and representative dataset is paramount to the success of the
system. Through comprehensive dataset collection methods, encompassing diverse
scenarios and environments, we ensure that the model is robust and capable of handling
real-world crowd management challenges effectively. Drawing inspiration from the
methodology of the Tensorflow Object Counting CSRNet, our approach follows a
structured framework encompassing various stages. Data preparation involves
preprocessing input data and generating density maps using Gaussian distribution
functions, laying the foundation for accurate crowd assessment. A custom data
generator class facilitates efficient data handling, ensuring smooth processing of
batches of images and corresponding density maps. The model architecture is
meticulously designed to leverage the strengths of deep learning, incorporating
convolutional layers for feature extraction, and reshaping to generate output heat maps.
During the training phase, a custom loss function is employed to compute the root mean
square error (RMSE), optimizing model performance. Training and validation are
conducted over multiple epochs, utilizing the custom data generator to enhance
efficiency. Following training, the model undergoes rigorous evaluation on sample
crowd images, with predictions compared against ground truth density maps to assess
its accuracy and efficacy. Visualization techniques are employed to provide qualitative
insights into the model's performance, facilitating a deeper understanding of its
capabilities. Finally, this study presents a comprehensive approach to robotic crowd
management through the integration of deep learning and multi-modal fusion
techniques.
