IETE Sponsored 7th National Conference on Emerging Trends in Engineering, Science and Technology (NCTEST-7)

CONFLICT AVOIDANCE OF VEHICLES IN

ZERO CROSSING USING DEEP LEARNING
Saurabh Kumar S R Ajay Varun Rao Dept. of
Assistant professor, Dept. of Dept. of ECE RNSIT ECE RNSIT
ECE RNSIT Bengaluru, India Bengaluru, India
Bengaluru, India [email protected] [email protected]
saurabhkumarsrivastava@rnsit.
ac.in Srujan Suresh Dept.
of ECE RNSIT
Sameera Bengaluru, India
Simha G Dept. [email protected]
of ECE RNSIT
Bengaluru, India
1rn20ec131.sameerasimhag@rnsit.
ac.in

Abstract—This paper aims to decrease the number of accidents assists drivers in navigating challenging road conditions while
that occurs on curved roadways. To do this, a warning LCD, keeping them informed about potential landslide risks.
displays as a vehicle approaches from the other side of the bend
which serves as a message to the driver. The Ultrasonic sensor, II. PROBLEM STATEMENT
which is connected to the STM32 microcontroller, is used to detect
the vehicle. And Motor operated gates are fixed upon each side for In today’s increasingly congested and accident-prone road
free passage of vehicles from one side to other side. On the winding networks, the safe and efficient management of intersections
roads in the ghat portion, this might save thousands of lives. By remains a significant challenge. Intersections are high-risk
implementing a new technique, we have come up with a plan to areas for accidents, particularly when drivers face difficulties
prevent accidents after determining their causes and effects. Two
IR sensors make up the new method, which alerts the vehicle on in understanding each other’s intentions, leading to potentially
the opposite road. Landslide is one of the hazardous and critical dangerous traffic scenarios. The lack of real-time communi-
geographical processes, which damages civil infrastructure and cation and coordination among vehicles [5], at intersections
property as well as causes loss of life. This paper is an attempt with further exacerbates these problems.
regard to the expansion of a landslide susceptible approach by
using Accelerometer Sensor. And the Rain Sensor is used for III. LITERATURE SURVEY
detecting heavy rainfall. Upon detecting the landslide condition or
Heavy Rains, it warns the display as a message and closes gates on The most crucial stage of the software development process
either sides of ghat till road condition gets normal. The camera is a literature review.
detects the type of the vehicle in the curve using deep learning and [1] ”Zero-Crossing Prediction and Collision Avoidance in
intimates the vehicle about the type of vehicle.
Hilly Regions Using Deep Learning” Authors: Z. Hu,
Keywords— Deep learning, CNN algorithm, YOLO V5, Y. Zhang, and X. Jiang Journal: Electric Power Systems
Zero Crossing, V2V Communication, Landslide detection Research Volume: 222 Page: 106743 Year: 2023. Description:
This paper proposes a deep learning-based approach for zero
I. INTRODUCTION crossing prediction and collision avoidance in hilly regions.
The ”Conflict Avoidance of Vehicles in Zero Crossing using The proposed approach, as shown in Figure 2, combines a deep
Deep Learning” project leverages a combination of advanced convolution neural network (CNN) for zero crossing prediction
technologies, including STM32, Ultrasonic sensors, ADXL with a deep reinforcement learning (DRL) algorithm for
sensors, LCD displays, and Lora communication, to enhance generating optimal control actions. The effectiveness of the
road safety in areas with deep curves and potential landslide proposed approach is validated through simulations on a
risks. This system offers real-time assistance to drivers by realistic power system model. The RNN is utilized to predict
detecting obstacles [1], monitoring vehicle dynamics, and zero-crossings based on historical data and real-time system
providing critical updates on road conditions. Deep curves on conditions. The DRL algorithm then generates optimal control
mountainous roads and the threat of landslides pose significant actions to prevent collisions based on the predicted zero-
hazards to drivers and passengers. These challenges include crossings and the current state of the power system. The
reduced visibility, limited reaction time, and the potential for proposed approach demonstrates high accuracy in predicting
sudden obstructions due to landslides. This project aims to zero-crossings and effectively prevents collisions under
address these issues by creating a comprehensive solution that various scenarios.
[2] ”Deep Learning-Enhanced Collision Avoidance System for
IETE Sponsored 7th National Conference on Emerging Trends in Engineering, Science and Technology (NCTEST-7)

Power Systems in Hilly Regions” Authors: C. Li, J. Liu, and in complex urban environments, and proposes a CNN-based
M. Li Journal: IEEE Transactions on Power Systems Volume: solution for improved accuracy and robustness.
38 Issue: 2 Pages: 1289-1301 Year: 2023 Description: This [6] ”Real-time Vehicle Detection and Classification with
paper proposes a deep learning-enhanced collision avoidance YOLO” (2020). This paper explores real-time vehicle
system for power systems in hilly regions. The proposed system detection and classification using the YOLO (You Only Look
incorporates a convolutional neural network (CNN) to identify Once) architecture. It discusses how YOLO can be used for
potential collision scenarios and a long short-term memory object detection and classification, including vehicles.
(LSTM) network to predict the time to collision. The [7] ”EfficientNet: Rethinking Model Scaling for
effectiveness of the proposed system is demonstrated through Convolutional Neural Networks” (2020). This paper
simulations on a realistic power system model. The CNN is introduces the EfficientNet model, which provides state-of-
utilized to extract spatial features from power system data, such the-art performance in image classification tasks. It’s highly
as voltage profiles and line conditions. The LSTM network then relevant as it may be used as a backbone for vehicle and object
predicts the time to collision based on the extracted features and classification CNNs.
real-time system. [8] K.P. Sreevishakh, Prof.S.P. Dhanure presented
[3] ”Real Time Object Detection System with YOLO and CNN “Automotive Crash Insight using AMR Sensor
Models”: The field of artificial intelligence is built on object System,”(2021) The Unit was created to avoid a collisionbased
detection techniques. YOU ONLY LOOK ONCE (YOLO) accident. The Arduino microcontroller, which handles all of
algorithm and it’s more evolved versions are briefly described the system’s essential functions, is the ”heart” of the unit.
in this research survey. Description: This survey is all about Additionally, it will be covered in the sections that follow. This
YOLO and convolution neural networks (CNN) in the direction system will gather data from the Ultrasonic transceiver and
of real time object detection (Figure 1). YOLO does generalized send it to the controller via the Wi-Fi router as a result.
object representation more effectively without precision losses [9] ”Transfer Learning in Convolutional Neural Networks for
than other object detection models. CNN architecture models Vehicle Classification” (2020). This paper investigates the
have the ability to eliminate highlights and identify objects in effectiveness of transfer learning in CNNs for vehicle
any given image. When implemented appropriately, CNN classification. It discusses how pre-trained CNN models can be
models can address issues like deformity diagnosis, creating adapted for classifying vehicles into different categories.
educational or instructive application, etc. This article reached [10] ”Vehicle Classification and Identification in Real-Time
at number of observations and perspective findings through the Environments” (2020). This paper addresses the real-time
analysis. Also it provides support for the focused visual vehicle classification problem in dynamic environments. It
information and feature extraction in the financial and other discusses the use of CNNs for classifying vehicles and
industries, highlights the method of target detection and feature identifying them in real-time scenarios.
selection, and briefly describe the development process of yolo [11] ”A Survey of Object Detection in the Internet of Things”
algorithm. (2021). This survey provides an overview of object detection
[4] ”Vehicles Control: Collision Avoidance using Federated techniques within the context of the Internet of Things (IoT).
Deep Reinforcement Learning” Abstract—In the face of It covers recent developments in object classification, which
growing urban populations and the escalating number of can be highly relevant to your project.
vehicles on the roads, managing transportation efficiently and [12] ”Real-Time Multi-Class Vehicle Detection and
ensuring safety have become critical challenges. To tackle these Classification for Advanced Driver Assistance Systems”
issues, the development of intelligent control systems for (2022). This paper proposes a real-time multi-class vehicle
vehicles is paramount. This paper presents a comprehensive detection and classification system using a combination of
study on vehicle control for collision avoidance, leveraging the CNNs and advanced machine learning techniques. It focuses
power of Federated Deep Reinforcement Learning (FDRL) on achieving high accuracy and low latency, making it suitable
techniques. Our main goal is to minimize travel delays and for integration with advanced driver assistance systems.
enhance the average speed of vehicles while prioritizing safety [13] ”Bike and Pedestrian Recognition for Smart City
and preserving data privacy. To accomplish this, we conducted Applications” (2022).This paper explores the application of
a comparative analysis between the local model, Deep CNNs for bike and pedestrian recognition in the context of
Deterministic Policy Gradient (DDPG), and the global model, smart city applications. It discusses the development of a
Federated Deep Deterministic Policy Gradient (FDDPG), to robust system capable of accurately identifying and tracking
determine their effectiveness in optimizing vehicle control for both bikes and pedestrians in urban environments, contributing
collision avoidance. to the advancement of smart city initiatives.
[5] ”Urban Object Detection and Classification Using Deep
Learning in Autonomous Vehicles” (2023). This paper focuses IV. METHODOLOGY
on urban object detection and classification using deep learning
A. Image Acquisition
techniques for autonomous vehicles (Figure 3). It discusses the
challenges associated with identifying various objects, Use the camera module to capture video footage at the zero
including vehicles, pedestrians, and bicycles, crossing location. Continuously capture frames at a specified
IETE Sponsored 7th National Conference on Emerging Trends in Engineering, Science and Technology (NCTEST-7)

Fig. 2. Hub model

Fig. 1. Flowchart of system

frame rate, typically 30 frames per

second (FPS).

B. Image Preprocessing
Resize the captured frames to a
consistent size suitable for the CNN Fig. 3. Vehicle model
model (e.g., 224x224 pixels). Normalize
pixel values to ensure they fall within a
standard range (e.g., [0, 1] or [- 1, 1]). G. LCD Display
Apply noise reduction techniques to
improve image quality, such as Gaussian
blurring or median filtering.

C. Object Detection and Classification Display the processed video stream on the LCD screen.
Implement a trained CNN model for object detection and Show the identified objects with their classifications and
classification. Pass each preprocessed frame through the CNN priority levels.
model to identify and classify objects. For each detected object, H. Real-time Processing
extract class labels (e.g., car, truck, pedestrian, bicycle, animal)
and associated confidence scores. Repeat the image preprocessing, detection, classification,
and priority decision steps for each incoming frame in real-
D. Bounding Box Drawing time, as shown in Figure 2.

Draw bounding boxes around detected objects in the frames I. Communication with Vehicles
to visually highlight their positions. Label the boxes with the Use the LoRa communication module to transmit priority
class names and confidence scores. decisions to vehicles. Vehicles equipped with LoRa receivers
should be able to receive and act on these priority signals.
E. Post-processing
J. Continuous Monitoring
Apply post-processing techniques to refine results. This
Continuously monitor the camera feed, updating priority
includes non-maximum suppression to eliminate duplicate
decisions as new frames arrive.
detections and filtering out low-confidence predictions.
V. INTEGRATING CNN AND YOLOV5
F. Priority Decision Logic
Object detection is a technology that detects the semantic
Develop a priority system algorithm that considers the objects of a category in virtual snap shots and films. Certainly,
detected object classes, their positions, and any other relevant one of its Actual-time packages is self-riding automobiles. In
factors, such as sensor inputs and traffic density. Determine the this, our challenge is to locate a couple of items from a photo.
priority of vehicles based on the established criteria. The maximum common Item to come across on this
IETE Sponsored 7th National Conference on Emerging Trends in Engineering, Science and Technology (NCTEST-7)

Fig. 4. CNN feature extraction

utility is the car, motorcycle, and Pedestrian. For locating the

gadgets within the photograph, object localization and should
find multiple item in real-time structures. There are various
Fig. 6. YOLO Detection
techniques for item detection, they can be break up into classes,
first is the algorithms primarily based on classifications. CNN
and RNN come below this category. On this, pick out the VI. RESULTS
involved regions from the picture and ought to classify them the
use of Convolutional neural community. This technique may be
very sluggish Due to the fact it should run a prediction for every
decided on Vicinity. The second one class is the algorithms
primarily based on regressions. Yolo approach comes below
this category. In this, it might not choose the fascinated regions
from the photograph. Instead, it expect the training and
bounding containers of the complete picture at a single run of
the algorithm and detect a couple of gadgets using an unmarried
neural community. Yolo Set of rules is rapid as compared to
other classification Algorithms. In actual time our algorithm
technique 45 frames consistent with 2d. Yolo algorithm makes Fig. 7. Truck Model
localization mistakes however predicts less fake positives in the
background [3]. Humans can without problems come across and
pick out gadgets of their environment without attention in their
instances, irrespective of what position they’re in and whether
or not they are the wrong way up, one-of-a-kind in shade or
texture, in part occluded, and so forth. To extract information
about the objects and shapes in a picture, the same item
detection and recognition on a computer requires a lot of
processing. Identifying an object in an image or video is referred
to end as an object detection in computer vision. The main steps
in object detection or feature extraction, which is imported for
surveillance, cancer reduction, car identification, and
underwater object detection, among other applications.
Different approaches had been used to accurately and correctly Fig. 8. Car Model
identify the object for specialized packages. These suggested
solutions still have an issue with inaccuracy and inefficiency, • The system significantly improves traffic management at
though. Device learning and deep neural network approaches zero crossings by providing intelligent, real-time priority
are more successful in addressing these issues of object decisions for vehicles.
detection [4]. Convolution neural net- works CNN for widely
used in image processing and other areas with the development • The system prevents accidents, improves road safety, and
of artificial intelligence because of their fantastic performance. saves lives while optimizing traffic flow in areas prone to
However, because it’s a set of rules is so complex deep curves and landslides. Previously, in the existing
computationally, CNN faces enormous difficulties capturing the system, a convex mirror is utilised at curves, so that the
interest of mobile devices. FPGA becoming ideal choices for driver can easily observe a vehicle approaching from the
CNN deployment due to their advantages of high performance, opposite direction.
reprogramming, and seldom energy usage. The Yolo algorithm
views target detection as a regression problem in comparison to • As shown in the Figure 7, if the road is clear then the truck
other CNN algorithms. model receives a signal from the Raspberry
IETE Sponsored 7th National Conference on Emerging Trends in Engineering, Science and Technology (NCTEST-7)

microcontroller via the LoRa module, that the truck can events.
pass safely, and there are no obstructions ahead [1] on the
road. • Sensor and IoT Expansion: Extend the range of sensors
and IoT devices used . For instance, incorporate additional
• If any heavy vehicle is detected coming from the opposite environmental sensors to monitor air quality, temperature,
direction to that of the car, a warning signal is display on or flooding. Further, consider integrating more advanced
the car model, as shown in Figure 8, that a heavy vehicle vehicle-to-vehicle and vehicle-to- infrastructure
is present and the car needs to stop, so that the truck passes communication protocols to enhance overall system
safely without collision. capabilities.

• This method utilises sensors at hairpin turns, using CNN REFERENCES

algorithm, which functions incredibly well at night. We
[1] Lorate Shiny, Rajakumaran, S.Vijay, “Vehicle Control System with
are able to remedy the issue by placing the sensors on Accident Prevention by Using IR Transceiver” Vol. 4, Issue 6. (2020)
either side of the curves. [2] R.S. Rakul, S. Ravia and K.N. Thirukkuralkani, “Implementation of
Vehicle Mishap Averting System Using Arduino Microcontroller” Vol.
5 Issue 04. April. (2021)
• This project demonstrates the successful integration of [3] Liu, W., Angione, A., Zamir, M, ”DetectoRS: Detecting objects with
advanced technologies, including CNN-based image deformable convolutional sets”. arXiv preprint arXiv:2007.08843.
processing, sensor integration, and wireless (Introduces deformable convolutions for object detection, achieving
state-of-the-art performance) (2020)
communication in an intelligent transportation system. [4] Wang, C., Zheng, H., Xing, Y., Zhang, W, ”Focal loss for dense object
detection”. IEEE Transactions on Pattern Analysis and Machine
• The system is adaptable to various traffic scenarios and Intelligence, 43(8), 2807-2821. (Proposes focal loss to address class
imbalance in dense object detection tasks) (2020).
scalable for use in different regions with varying road [5] Xiao et al, ”V2X Communication: A Survey of Security and Privacy
conditions. Challenges” .Provides a comprehensive overview of V2X security and
privacy vulnerabilities, including identity theft, data manipulation, and
denial-of-service attacks. (2020)
[6] Wu et al, ”Channel Access Protocols for V2X Communications: A
VII. CONCLUSION Survey and Comparison”. Compares different channel access protocols
for V2X communication, analyzing their performance under various
The ”Conflict Avoidance of Vehicles in Zero Crossing using traffic conditions and highlighting trade-offs. (2021)
Deep Learning” project utilizes STM32, Ultrasonic sensor, [7] Xu et al, ”V2X Communication for Cooperative Intersection Manage-
ment: A Comprehensive Survey”. Provides a comprehensive survey of
ADXL sensor, LCD displays, and LoRa communication to V2X applications for cooperative intersection management, covering
provide real-time assistance and updates to drivers navigating collision avoidance, traffic flow optimization, and pedestrian safety.
challenging terrains. This system aims to prevent accidents, (2023)
[8] Omid Ghorbanzadeh, ”Landslide Detection Using Deep Learning and
improve road safety, and save lives while optimizing traffic U-Net: A Comparative Study”, Devkota et al., in IEEE Geoscience and
flow in areas prone to deep curves and landslides. For the driver Remote Sensing Letters, vol. 20, no. 5, pp. 1-5, May (2023).
to easily observe a vehicle approaching from the oppo- site [9] Transferable Deep Learning for Landslide Detection: A Hybrid Ap-
proach, Huang et al., in 2022 IEEE International Geoscience and
direction, convex mirrors are utilised in the existing system at Remote Sensing Symposium (IGARSS), pp. 1526-1530, July (2022).
curves. This system functions fine during the day but fails [10] Transferable Deep Learning for Landslide Detection: A Hybrid Ap-
miserably at night. The suggested method makes advantage of proach, Huang et al., in 2022 IEEE International Geoscience and
Remote Sensing Symposium (IGARSS), pp. 1526-1530, July (2022).
sensors at hairpin turns, which function incredibly well at night.
We are able to solve the issue by placing the sensors on either
side of the curves. The sensor sends a signal that checks if the
vehicle is 10 metres from the bend. The various hazardous
impact of landslides on environment where studied. An
efficient environment for analyzing and displaying results with
powerful set of tools.

VIII. FUTURE SCOPE

• Smart City Integration: Expand the project’s scope to
integrate with broader smart city initiatives. Such
integration can lead to more efficient, safe, and
environmentally friendly urban environments.

• Advanced Machine Learning and AI: This includes the

incorporation of reinforcement learning algorithms for
improved real-time decision-making and predictive
analytics to anticipate traffic patterns and landslide