Institutional Digital Repository (Indian Institute of Technology Kharagpu)

Sankhla, Sangeeta

Establishing Environment-Friendly Approaches for Cellulose Nanofibers Extraction from Secondary Bioresources and Its Various Applications

Das, Gargi

Fault-Tolerant Unwinding-Free Satellite Tracking Control in the Framework of Modified Rodrigues Parameters

Agarwal, Ankit

Retrieving Hidden Physics of Interfacial Interactions via Physics Informed Neural Network Based Modeling in Nanofluidics

Vaidya, Pravin Vasudeo

Net Shape Forming of Dense and Porous Structures for Customized Ti6Al4V Implant Through Slurry Processing

Samal, Bijaya Bikram

Four-Dimensional Printing: Shape Programming, Material Characterization, and Parametric Optimization

Yadav, Rahul

Role of Tailored Residual Stress on Control of Burr Formation in Micro End Milling: Experiments and Numerical Modelling

Sambangi, Ramesh

Title: A Machine Learning Approach for Network-on-Chip Architecture Design
ABSTRACT
Escalating computational demand, specially due to machine learning and high performance
computing (HPC) workloads has made the design of multiprocessor system-on-chip (MPSoC)
complex and challenging. These systems necessitate parallel processing, immense computational
power, and high-speed on-chip communication infrastructure. While computational needs can be
taken care of by integrating large number of processing elements into a chip, effective
communication between them becomes a challenge. Network-on-Chip (NoC) technology
addresses the issue of energy-efficient and high-performance on-chip communication. In this
direction, efficient mapping of an application onto NoC is a pertinent step that remains a
challenge, demanding improved algorithms to overcome the limitations of existing heuristic
approaches that often encounter local minima.
This dissertation focuses on optimizing the application mapping process with packet latency and
communication cost as the key performance metrics. Traditional mapping methods need to
invoke NoC simulators for latency estimation, incurring high execution time for the mapping
process. To address this issue, we propose a LPNet model, a deep learning-based latency
prediction model. Based on the simulation results, it has been observed that LPNet reduces errors
when compared with the simulator results. LPNet has been integrated into a Discrete Particle
Swarm Optimization (DPSO) algorithm to identify superior mapping solutions and reduced
optimization time.
ATSR (Active Search), a machine learning-based mapping algorithm has been specifically
designed for application mapping in 2D mesh NoCs. Machine Learning (ML) algorithms have
gained significant importance in the field of combinatorial optimization due to their ability to
effectively tackle complex problems, compared to the traditional heuristic approaches. ATSR
harnesses the power of message passing neural networks and pointer networks, allowing it to
navigate complex mapping spaces efficiently. To enhance its capabilities further, a new IP-core
numbering scheme has been introduced that improves model accuracy and compatibility across
various NoC sizes. Simulation has been performed to validate and assess the practical
applicability of the ATSR algorithm in real-world scenarios.
Another scheme, named NCTPAM (Neural Congestion-aware Through-silicon via (TSV)
Placement and Application Mapping), an ML based approach for application mapping and TSV
placement in partially connected 3D mesh NoC has been introduced. Finding optimal mapping
solutions and TSV placements is an NP-hard problem. NCTPAM incorporates graph attention
networks and pointer networks in its architecture. A method has been proposed to balance the
load across vertical connections, thereby improving overall reliability. Performance of these
proposed strategies has been studied analytically and also via simulation.
Keywords: Network-on-chip (NoC), NoC analytical model, application mapping, machine
learning, message passing neural network (MPNN), pointer networks, graph attention networks,
reinforcement learning, sequence-to-sequence models.


A Machine Learning Approach for Network-on-Chip Architecture Design

Halder, Joyanti

Thesis title: 
Ni1-xCoxFe2O4 (0.0 ≤ x ≤ 1.0) Based Nanostructures for High Performance Supercapacitors  
ABSTRACT 
Mixed metal oxides are gradually gaining prominence as electrode materials for 
supercapacitors. Such mixed metal oxides, consisting of more than one transition metals, allow 
synergistic contributions from each redox element in the material. This brings significant 
enhancement in the pseudocapacitance value. The rich redox chemistry, easy synthesis process, 
cost-effectiveness and environmentally benign nature of these mixed metal oxides are other 
advantages, useful for supercapacitor applications. Amongst the different mixed metal oxides, 
inverse spinel ferrites are being extensively investigated. This thesis deals with the use of 
NiFe2O4 and Ni1-xCoxFe2O4 (0.0 ≤ x ≤ 1.0) as efficient electrode materials for pseudocapacitors. 
NiFe2O4 has an inverse spinel structure, where Ni
2+ occupies the octahedral sites while Fe3+ 
favours both the octahedral and tetrahedral sites according to the law of minimization of the 
crystal field stabilization energy (CFSE). The structure allows the hopping of electrons, a 
phenomenon that makes them excellent for electrochemical activity.  
 Out of the different compositions, it is shown that the Ni0.5Co0.5Fe2O4 has the best 
performance. The reasons driving a high electrochemical performance in this composition are 
discussed in detail. A new theoretical model is also proposed to explain the change of the 
diffusive nature of the electrolyte ions across the SEI, which leads to high redox activity in the 
materials reported in the thesis. To further increase the specific capacity value, a strategy of 
using redox additive modified electrolytes is proposed. The fruitfulness of the strategy is 
established by using K3[Fe(CN)6]and KI as redox additives in pure 3 M KOH electrolyte. The 
discharge capacity jumps from ~ 213 F g-1 to ~ 476 F g-1 when the pristine 3 M KOH electrolyte 
is modified with optimized concentration of K3[Fe(CN)6]. For the case of KI modified 
electrolyte, the value of specific capacitance increases from ~ 213 F g-1 to ~ 260 F g-1. 
Interestingly, the electrochemical performance of all the synthesized materials (i.e. α-Fe2O3, 
NiO and Ni1-xCoxFe2O4 (0.0 ≤ x ≤ 1.0)) can be tuned by varying the external (applied) magnetic 
fields. The usefulness of these results for industrial applications are also presented in the thesis. 
 
Keywords: Supercapacitors, Pseudocapacitors, Metal oxides, Inverse Spinel Ferrite, Magnetic 
Supercapacitors, Redox Additives. 
 


Ni1-xCoxFe2O4 (0.0 ≤ x ≤ 1.0) Based Nanostructures for High Performance Supercapacitors

Pramanik, Anima

LEARNING ALGORITHMS FOR EVENTS DETECTION
FROM TRAFFIC SURVEILLANCE VIDEOS
Thesis submitted in the partial fulfillment
of the requirements for the Degree of
Doctor of Philosophy (Ph.D.)
by
Anima Pramanik
(17IM91P01)
Under the supervision of
Prof. J. Maiti
Department of Industrial and Systems Engineering, IIT Kharagpur
and
Prof. Sankar K. Pal
Center for Soft Computing Research, ISI Kolkata
Department of Industrial & Systems Engineering
Indian Institute of Technology Kharagpur
June, 2024
©Anima Pramanik. All Rights Reserved
Abstract
Name of the student: Anima Pramanik Roll No: 17IM91P01
Degree for which submitted: Doctor of Philosophy (Ph.D.)
Department: Department of Industrial & Systems Engineering
Thesis title: LEARNING ALGORITHMS FOR EVENTS DETECTION
FROM TRAFFIC SURVEILLANCE VIDEOS
Thesis supervisor(s): Prof. J. Maiti, and Prof. Sankar K. Pal
Month and year of thesis submission: June, 2024
The usefulness of computer vision (CV) in detecting events has been established in diverse
domains of application (i.e., military, security, life, and medical). However, its exploration
in detecting real-time traffic events in road transportation domain is limited. This necessi-
tates the investigation of CV algorithms in traffic safety domain. This research aims at the
development of various learning algorithms for detecting the real-time traffic anomalous
events (i.e., pre-events, current-events, and post-events) and taking the on-time corrective
actions in order to enhance the traffic safety. The thesis addresses four objectives as: to
develop various learning algorithms for (i) both foregrounds and multi-class objects detec-
tion and tracking, (ii) vehicle-related traffic pre-events detection, (iii) traffic current-events
detection, and (iv) person-linked traffic anomalies detection.
The first objective addresses the use of granular computing in foregrounds and multi-class
objects detection and tracking. Based on the object of interest used for traffic anomaly
detection, both unsupervised and supervised algorithms are developed for foregrounds and
multi-class objects detection and tracking, respectively. Here, foregrounds detection and
tracking (based on unsupervised learning) are done using the spatio-neighborhood tempo-
ral granulation, and distance- and area-based data association, respectively. On the other
hand, two supervised algorithms, namely granulated RCNN (G-RCNN) and multi-class
ii
iii
deep SORT (MCD-SORT) are developed for multi-class objects detection and tracking,
respectively. Efficacy of both of the unsupervised and supervised algorithms for fore-
grounds and multi-class detection and tracking are demonstrated over various benchmark
data (PASCAL VOC, Urban Tracker, YouTube8M, and MOT15) and one real-life traffic
data, namely Video Analytics Road 2019 (VAR19). Moreover, superiority of the developed
algorithms over some recent state-of-the-art is also proved.
The second objective comprises of two sub-objectives: (i) development of a conceptual
framework for video surveillance system in enhancing the traffic safety management and
(ii) development of various unsupervised algorithms for the detection of vehicle-related
traffic pre-events, such as speed violation, one-way/oncoming traffic, overtaking, illegal
parking, roundabout violation, and wrong drop off location of passengers. The developed
unsupervised algorithms for the detection of aforesaid traffic pre-events are modeled using
various rules based on some defined spatial and temporal features. The efficacy of aforesaid
algorithms is demonstrated over 80 real-time traffic videos containing both normal and
abnormal (i.e., traffic pre-events) scenarios. A comparative study is also conducted to
prove the superiority of developed algorithms for aforesaid traffic pre-events detection.
The third objective addresses the use of spatio-temporal features (STFs), rough fuzzy gran-
ulation, and Z-numbers for handling the uncertainty issue arising in between two traffic
current-events, namely collision and near-miss. Various salient STFs are defined corre-
sponding to each detected foreground region and used for obtaining the crisp granules
over the video frames. These granules represent the approximate anomaly-prone regions.
As there is no definite class (collision, near-miss, and/or normal scenario) assigned to these
regions, fuzziness or uncertainty may inherently present in these granules, thereby result-
ing in fuzzy information granules. The said uncertainty has been modeled by defining
two rough fuzzy granules (RFGs) (one is for collision event and another is for near-miss
event) along with their roughness scores. For each RFG, Z-number is computed based
on the membership value of its roughness score to ensure a higher degree of reliability in
the detection of anomaly/event-class. All these characteristics constitute an unsupervised
algorithm, namely Z-numbers-based spatio-temporal rough fuzzy granulation (Z-STRFG).
The efficacy of Z-STRFG has been demonstrated over 130 real-time traffic videos contain-
ing collisions, near-misses, and normal traffics. The superiority of Z-STRFG over some
iv
state-of-the-art is also proved through extensive experiments.
In the fourth objective, two semi-supervised algorithms, namely Z-numbers-based spatio-
temporal granulation (Z-STG) and deep transfer learning (Z-DTL) are developed for the
detection of four types of person-linked anomalies (PLAs) which are categorized into three
classes: (i) pre-events (i.e., causes: (pedestrian crossing violation (PCV) and pedestrian
proximity to moving vehicle (PPMV)) of current-events), (ii) post-event (i.e., consequence
(pedestrian/driver fall (PF)) of current-events), and (iii) personal protective equipment
violation (PPE-V) (i.e., violation of using the hard-hat which can be responsible for mit-
igating the impact of consequence). Z-STG addresses the issue of detecting the occluded
or blurred pedestrian crossing as well as its violation. Z-STG consists of two steps. In
the first step, various spatio-color and temporal features are defined and used for ex-
tracting the pedestrian crossings in terms of crisp granules. Whereas, in the second step,
object-level features are fused with the detected pedestrian crossing features for generat-
ing various rules in order to detect the PCV and PPMV. On the other hand, Z-DTL for
fall and PPE-V detection deals with the issues of low illumination, occlusion, and varying
behavior of the traffic events from video to video. The Z-DTL incorporates a new deep
network, namely G-RCNN3 with a two-dimensional bidirectional long short-term memory
(Bi-LSTM) using transfer learning. G-RCNN3 is an advanced version of G-RCNN. The
output layer (FC3) of G-RCNN3 contains three dimensional feature maps. Each dimen-
sion represents the feature map corresponding to either person, fall-event, or hard-hat
(PPE). The 1st- and 3rd-dimensional feature maps are used for re-training of G-RCNN3.
Whereas, the 1st- and 2nd-dimensional feature maps are used for training and finetuning of
Bi-LSTM. In both Z-STG and Z-DTL, Z-number is computed to ensure the reliability of
the detected PLA-class. The superiority of Z-STG and Z-DTL over some state-of-the-art
is also established through extensive experiments.
The thesis has made contributions from both theoretical development and practical im-
plication points of view. From the theoretical development point of view, the thesis con-
tributes as follows: (i) an unsupervised algorithm is developed for foregrounds detection
and tracking using the commonality between spatio-color contour-based neighborhood
and temporal granules, and distance- and area-based data association, respectively, (ii) a
v
supervised detector, namely G-RCNN is developed for multi-class objects detection by in-
corporating the concept of spatio-temporal granulation within deep CNN, (iii) a supervised
tracker, namely MCD-SORT is developed for multi-class tracking by restricting the data
association between detected object(s) and track-lets of same categories, (iv) various unsu-
pervised algorithms are developed for detecting vehicle-related real-time traffic pre-events,
(v) an unsupervised algorithm, namely Z-STRFG is developed for real-time traffic current-
events detection by handling the uncertainty issue arising between collisions, near-misses,
and normal scenarios, (vi) a semi-supervised algorithm, namely Z-STG is developed for
detecting person-linked traffic pre-events (i.e., PCV and PPMV) using spatio-temporal
granulation, rule base system, and Z-numbers, (vii) another semi-supervised algorithm,
namely Z-DTL is developed for detecting two person-linked anomalies which are conse-
quence (i.e., fall) of current-events and violation of using the mitigating factor (i.e., PPE-
V), (viii) the research demonstrates a way of incorporating the recently emerged granular
computing in deep learning network for on-line object tracking and classification directly
from the input videos, and (ix) the effectiveness of granulation in detecting various traffic
pre-events and current-events, is demonstrated in this research. On the other hand, from
practical implication point of view, the thesis contributes as follows: (i) developed algo-
rithms can automatically detect anomalous events from 24 × 7 live stream traffic videos,
(ii) after detection, a video summary of the anomalous event can be generated automat-
ically and used for future training, (iii) reduction of human effort in order to search the
traffic anomalous events from huge amount of videos, and (iv) after detection, an alarm
signal can be sent automatically to the traffic control room for taking on-time corrective
actions in order to either stop re-occuring the current-events (i.e., collision and near-miss)
or mitigate the impact of consequence for same current-events, thereby enhancing the
traffic safety.
From the analyses presented in the thesis, some key findings are obtained. They are:
(i) commonality between spatio-color contour-based neighborhood and temporal granules
represents the foreground regions in videos more effectively as compared to spatio-color and
temporal granules, (ii) region of interest (RoI)-pixels detected by the developed G-RCNN
are more representative of object classes as compared to features generated by RCNN and
its variants, (iii) comparative results signify the relevance of spatio-temporal information
vi
over spatial information for the task of video object localization, (iv) G-RCNN is found
to be the best among state-of-the-art video object detectors with mean average precision
(mAP) of 80.6% on the generated data, namely VAR19, (v) among all the combinations of
supervised detectors and trackers considered in the experiment, G-RCNN + MCD-SORT
is found to be the best, (vi) algorithms for the detection of vehicle-related traffic pre-
events and scene-based traffic current-events are modeled using foreground-level features;
therefore, foregrounds detection and tracking is the prior task in such cases, (vii) person-
linked traffic anomalies are related to a specific object class (i.e., person); therefore, in
such cases, multi-class objects detection and tracking is the prior task, (viii) the developed
conceptual framework for traffic safety management is generic in nature and can be used in
any road context to detect vehicle-related traffic pre-events, (ix) for traffic current-events
detection, the use of RFGs and Z-numbers in handling the uncertainty issue is more
effective than using only salient cost functions, (x) the use of Z-numbers for each RFG in
ensuring the reliability of the detected current-event (collision and near-miss) is unique,
(xi) Z-STRFG is found to be the best among state-of-the-art video summarization and deep
CNN-based traffic anomaly detection techniques with accuracy of 96.8% over a new traffic
data, namely Anomaly 2020 (An20), (xii) in Z-STG, the incorporation of Z-numbers with
spatio-temporal granulation is found to be very effective in detecting blurred/occluded
pedestrian crossing and its violation (i.e., PCV and PPMV), (xiii) in Z-DTL, the fusion
of G-RCNN3 with Bi-LSTM using transfer learning and Z-numbers shows an adequate
performance in PF and PPE-V detection, and finally, (xiv) experimental results prove
the effectiveness of Z-STG and Z-DTL over some state-of-the-art for person-linked traffic
anomaly detection.
Keywords: Traffic safety management, Computer vision, Multi-object detection and
tracking, Vehicle-related traffic pre-events detection, Traffic current-events detection, Person-
linked traffic anomaly detection, Granulation, Rough fuzzy inference system, Uncertainty
handling, and Rule-base system.


Learning Algorithms for Events Detection from Traffic Surveillance Videos

Samanta, Sounak

Algorithms and Architectures for Estimation of Radar Pulse Parameters in Modern Electronic Support Receivers

Not a member yet