Self Navigating Car
Our project is about making a fully autonomous car. We?re incorporating the two most important features of a robotic vehicle: lane navigation and object identification and classification. Another feature that will be incorporated in our vehicle is that it will be able to see through fog with t
2025-06-28 16:34:57 - Adil Khan
Self Navigating Car
Project Area of Specialization Artificial IntelligenceProject SummaryOur project is about making a fully autonomous car. We’re incorporating the two most important features of a robotic vehicle: lane navigation and object identification and classification. Another feature that will be incorporated in our vehicle is that it will be able to see through fog with the help of a thermal camera. Up till now, there haven’t been any implantation of fog detection feature on large-scale.
The accuracy and reaction time of the lane detection and object identification features in a driverless car are of utmost important because the slightest inaccuracy or a delay in the reaction of the car can cause accidents. Thus, for our project, we researched on all the methods used for these two features. We then researched on the best technologies and the most accurate algorithms that could be used to minimize the reaction time and reduce the number of accidents.
We'll use sensor technologies to navigate the car safely on road. Least amount of time will be spent in hardware assembling and we will focus on the machine learning algorithms and techniques to be applied in the design of an autonomous driving system on these two perception features. Our prototype will process data in real-time and hence our main area of research is to ensure a deep learning model that is quick, highly efficient and allows our car to work in unknown environments. Our prototpe can can have generic applications and be tailored for specific needs
Project ObjectivesAutonomous vehicles are already driving themselves in test mode down the streets in the US. But their onboard navigation systems still can’t help them maneuver safely through heavy or even light fog. This project aims to make an autonomous car which will also be able to drive in foggy conditions.
Our target is also to come up with a highly efficient model to allow processing time to reduce substantially, hence allowing our car to speed up and work in unknown environments. We're targeting the best algorithms for both the robotic features of our car i.e lane detection and object classification and identification
About the first feature of the car i.e. lane detection, the previous proposed systems use the traditional Hough transform is used which is a rather inefficient model when vehicle must run on unknown terrains with no lane edges. Our objective is to use an accurate method so that it can drive without steering offsets and in unknown conditions, i.e. roads with faded lane marks.
About the second feature of the car i.e. object detection and classification, our objective is to reduce the processing time when the car detects an object, because the slightest delay can be dangerous. Our protoype car will able to detect miniature sized objects and react accordingly. It will stop for pedestrians, stop signs & red traffic signal, speed up at green signals, slow down or speed up at speed limit signs.
The plan will be implemented on a small and simple autonomous vehicle starter kit, so that least time is spent in assembling the hardware. The car kit is a complete prototype of a vehicle, and it includes all the necessary parts such as wheels, motors, motor shield, chassis etc.
A camera will be used along with this vehicle to capture 2D images, which will make the car ‘see’ what’s in its surroundings, whereas ultrasonic sensors and LiDAR will help the vehicle to make a 3D environment and detect the distance of the objects in the surroundings. An embedded processor, NVIDIA Jetson Nano, will be integrated with the vehicle. This will be the brain of the car and will make it decide what to do after it gets the data from the camera and sensors. Our deep learning models will be interfaced with this processor.
The vehicle needs a large amount of data on which it would be trained. The data needed for this will be collected manually. We’ll write a remote-control program to remotely steer the vehicle. We’ll then drive it on several manually designed tracks and have it save down the video frame as well as the car’s steering angles at each frame. This is the best way since it would be simulating a real person’s driving behavior.
Miniature traffic signs and small toy figures will be used to train for object detection with the help of single-shot multibox object detection algorithm.
A vulnerability in object detection happens during foggy conditions. Conventional cameras are unable to identify and detect objects and thus the reaction time under these conditions increase significantly. A way to tackle this issue is through thermal imaging. This is a new area of research and thus related technology is expensive. We will therefore be using a low-resolution (less expensive) thermal camera to achieve better reaction times in comparison to conventional cameras. The dataset for foggy conditions will be collected from the internet and our model will be run and tested on it. If need be, we will have to generate our own dataset using a fog machine and our thermal camera.
Benefits of the ProjectMore than 90% of fatal road accidents are caused due to human error. Autonomous cars have the potential to reduce deaths and injuries from car crashes, particularly those that result from driver distraction. However, even self-driving cars’ slow reaction time can cause accidents. Our project will overcome this hurdle by integrating the most accurate and fastest algorithms in the vehicle.
Secondly, the lane detection feature will allow the autonomous vehicle to constantly monitor surrounding traffic and respond with finely tuned braking and acceleration adjustments. This can enable the AV to travel safely at higher speeds and with reduced headway (space) between each vehicle.
If implemented on a large scale, this project can prove to be a milestone towards reducing the aggressive traffic conditions.
Moreover, the biggest current problem for autonomous cars is that they are unable to react under harsh conditions which can include, but is not limited to, fog. Foggy weather which impact driving for cars on public roadways is what major car manufacturing companies are trying to address. Our project will help understand how fog behaves and what method should be taken in order for images to be processed through a thermal camera and the car's reaction will be taken into consideration. The ability of our vehicle to see and drive through fog safely will be the distinct feature of as it will expand the scenarios in which the vehicle can drive autonomously.
Even though our model will be run on a small scale and most of the data fed will be of miniature objects, the entire model will still be robust enough to be implemented on a large scale. The model will have to be tweaked and tested and will be able to provide consumers a chance to enjoy all hands off the wheel, especially under fog conditions where even human eye fails to see.
The project will open plenty of areas to research for us including deep learning, image processing, hardware interfacing etc. Perhaps the most crucial benefit from this project would be taking a step forward in being able to navigate through unknown environments.
Technical Details of Final DeliverableAll programming for the processor, NVIDIA Jetson Nano, will be done on Python.
The vehicle will learn lane detection and object identification through machine learning algorithms for which we’ll use the SSD (Single-Shot Detector) object detection model as it is most accurate and fastest algorithm up till now.
The camera installed on the vehicle will give live video of the environment, which is basically a set of pictures. OpenCV, which is a powerful open source library for image processing, will be used to assess the pictures so that the vehicle can make sense of what it is seeing. Numpy and Matplotlib are two very useful python libraries that we will use in conjunction with OpenCV for image processing and rendering.
To apply the machine learning algorithms, we will use Google’s TensorFlow library as it is currently the most popular python library for Deep Learning. It can be used for image recognition, face detection, natural language processing, and many other applications.
COCO (Common Object in Context) object detection model will be used for transfer learning. Using pre-trained model will make our data collection process smaller.
The night vision camera, Pi NoIR Camera V2, will be used to capture videos. The TF-Luna LiDAR from SmartFly company is a single-point ranging Lidar, based on Time of Flight principle and it will provide the vehicle with distance information of the surrounding objects.
Our final deliverable will then be a prototype car which can easily navigate through unknown terrains and be able to react to certain objects in real-time. In case of fog, the car should be able to maneuver just as well.
Final Deliverable of the Project HW/SW integrated systemCore Industry TransportationOther Industries Energy Core Technology Artificial Intelligence(AI)Other Technologies RoboticsSustainable Development Goals Industry, Innovation and InfrastructureRequired Resources| Item Name | Type | No. of Units | Per Unit Cost (in Rs) | Total (in Rs) |
|---|---|---|---|---|
| Total in (Rs) | 80000 | |||
| RC Car Kit | Equipment | 1 | 5000 | 5000 |
| NVIDIA Jetson nano dev kit | Equipment | 1 | 20000 | 20000 |
| Lepton 1.5 (Thermal Camera) | Equipment | 1 | 21000 | 21000 |
| LiDar Sensor | Equipment | 1 | 9000 | 9000 |
| Miniature Objects | Miscellaneous | 10 | 100 | 1000 |
| Li-ion cells | Miscellaneous | 4 | 200 | 800 |
| Pi NoIR Camera v2 | Equipment | 1 | 5000 | 5000 |
| Fog Machine | Equipment | 1 | 10000 | 10000 |
| Markers | Miscellaneous | 2 | 100 | 200 |
| Shipment costs | Miscellaneous | 1 | 8000 | 8000 |