Autonomous vehicle using Computer Vision

Google Drive Link; Methodology pipeline, Results, video of my ongoing work, and so on. Everything is on this Link. It will better guide you about our progress. Autonomous driving a car that drives on its ow

Project Title

Autonomous vehicle using Computer Vision

Project Area of Specialization

Artificial Intelligence

Project Summary

Google Drive Link; Methodology pipeline, Results, video of my ongoing work, and so on. Everything is on this Link. It will better guide you about our progress.

Autonomous driving a car that drives on its own, what will traffic look like in a few years when will we see the self-driving car on the streets and how is it going to work. Autonomous cars are the future smart cars anticipated to be driverless, efficient, and crash avoiding ideal urban car of the future. To reach this goal automakers have started working in this area to realize the potential and solve the challenges currently in this area to reach the expected outcome. Today, I want to share some insights into one part of the software, that enables our vehicle to drive autonomously known as the Computer Vision System. This module is responsible for understanding the world around our vehicle based on the real-time stream from Cameras mounted all-around the vehicle. Given the camera streams, we compute several important outputs on each image using neural networks. That includes computing, for each object of interest in the scene, a 20 box that encloses it, and along with a mask indicating, which pixels belong to this object a task that is called instance segmentation. We also compute what semantic class each pixel belongs to. For example, is it pedestrian? A sidewalk? Car? And we can also predict the depth of each pixel in the real world, even though cameras can’t measure distance directly. In addition to this, we also predict the 3d bounding box location for each dynamic object of interest in the scene, just from cameras. There are many objects that we see in the real world that our vehicles are able to detect and classify. For Instance, pedestrians, pedestrians with different appearances, and poses, pedestrians on scooters, dogs, motorcycles, vehicles, and many more. In addition to this classification, we also compute several attributes for our detected classes that are useful for the rest of the stack. Here I consider describing some of the attributes for just two of our classes; pedestrians and vehicles. First, we have Skelton detection, for each pedestrian, in object detection, we detect the position of all the key points on their skeletons. These key points can then be used for the higher-level tasks, such as tracking, gesture detection, prediction of intent. We can also get a clue whether the pedestrian stops or is in motion by referring to a bounding box with some clue images, which gives clue to the car whether a person is crossing the road or not. In our software, we have a network that is able to produce these signals, which we also can visualize using Machine Human Interface.

Project Objectives

Google Drive Link; Methodology pipeline, Results, video of my ongoing work, and so on. Everything is in this Link. It will better guide you about our progress.

Nowadays, safety is a priority and challenge for technology developers whenever a new technology is going to introduce. You can take an example of a recent work of the autonomous car. They are working their best to make a self-driving car that ensures human safety and avoids incidents that happen because of human mistakes. Approximately 1.35 million people each year die in road crashes. On average, roughly 3,700 people lose their lives every day on the highways. A further 20-50 million endure crucial injuries, often culminating in everlasting disabilities. We can easily avoid all road accidents by implementing deep learning and machine learning to prepare improved driving assistant, which can warn drivers about the danger and save a certain number of lives from crashes. 

Between 2018 and 2019, 10779 accidents were reported in Pakistan and 4878 of them were fatal. These accidents took 5932 lives. The statistic shows that driving a car manually has proven to be fatal and hazardous. It becomes very difficult for a person to manage all-day stress and then drive home safely. Normal human mind thousands of thoughts and a frequent-rare scenario comes where he/she cannot concentrate on driving. Once our research is completed, we will be able to help every driver in real-time. The AI algorithms of ours will be able to detect lane lines for the driver, keep the car in the center of the lane, use the right speed, give recommendations when a traffic sign is detected, and provide a smooth and safe driving experience. Our key objective of self-driving cars is to ensure road safety from the precarious accidents of people.

Our other objective is to acquire accuracy and precision in Autonomous cars so that we can avoid pernicious accidents. Accuracy and efficiency are two main problems when it comes to lane line detection. Engineers are spending their days and nights designing more accurate and more efficient algorithms for lane line detection. We aim to design such an algorithm that will be able to detect lanes and draw lines on them with almost 100% accuracy. Our program will be able to know the slightest change in curves of the lane. Object detection requires a large number of datasets to be fed. A self-driving car must know about its surroundings. If the objects are not detected, we cannot make a self-driving car. Our aim is to make a car smarter by allowing it to detect almost every relevant object that surrounds it.

Project Implementation Method

Google Drive Link; Methodology pipeline, Results, video of my ongoing work, and so on. Everything is on this Link. It will better guide you about our progress.

For Autonomous Car, the complex and challenging task of this Advanced Driver Assistance to achieve the desired safety on roads and danger alerts are Lane Line or Boundaries Detection (white and yellow lines on the road), Obstacles detections (cars, pedestrians, trees, etc.), Traffic sign (Standard signs) and Traffic Lights. Here in this project, we will take real-time video streaming by using a camera to recognize the multiple Objects (moving and stationary) and Lane lines on roads. Hence, car handling and speed would control accordingly. The camera will mount on the car, one camera will be on the top of the vehicle, the second one will be the back camera, and the other two cameras will mount on the left and right side of the car, and this is how the real-time streaming would take. The different algorithms at the backend using distinct Open sources will work on these streaming. It will extract the data, in the format of an array, from the image, as real-time streaming comprises the sequence of image frames. It will locate our Region of Interest (RoI), and hence everything will work and control accordingly. 

Uses lane line detection method to detect objects for an autonomous vehicle. The main algorithm for lane line detection is Hough Transformation and its accuracy is above 90%. The object tracking in this research is done using the tool TensorFlow. The algorithm receives the identified object and processes it. The researchers used self-generated image sequences to test the program and its accuracy was 90.88%.

Benefits of the Project

https://drive.google.com/drive/folders/1pX6Lgj52tojdoizliY-O8GJPThZz1gCT?usp=sharing

1)  90% reduction in traffic deaths. Yes, you read that right.

2)  60% drop in harmful emissions.

3) Eliminate stop-and-go waves by 100% ...

4) 10% improvement in fuel economy.

5) 500% increase in lane capacity.

6) 40% reduction in travel time.

7) Consumer savings of billions.

8) Transportation Accessibility.

9) Reduce Transportation costs.

Technical Details of Final Deliverable

https://drive.google.com/drive/folders/1pX6Lgj52tojdoizliY-O8GJPThZz1gCT?usp=sharing

Object detection using Convolution Neural Networks (CNN) are comprised of three types of layers. These are convolutional layers, pooling layers, and fully-connected layers. When these layers are stacked, a CNN architecture has been formed.

Once we know how to do classification, it is easy to extend the network to do Localization. In classification, the last FC layer outputs 1 value per class. But for Localization, we need 4 outputs per class. These 4 outputs would correspond to h, w, x0, y0 coordinates of the bounding box. We just need to train the network to output correct values for these co-ordinates. The difference from the classification is that there would be no Softmax layer. In order to localize the object, we need to find the error rate between the expected bounding box to the ground-truth bounding box, for this purpose, we use the Pseudo Hueber loss this is the best of both losses if a data point has a relatively low error, we take the L2 loss if the data point is an outlier, we take the absolute loss. This way, we come up with a highly accurate bounding box of the object in an input image. Finally, combining the results of both classifier and Bounding Box Regressor, we infer the type of the object and its location. Note that, for the same object, if there are 10 classes in your dataset, 10 different classification scores, and Bounding Box coordinates will be output by the network. But we only pick the one which has the highest confidence score in the classification output

By implementing the fully connected layers as convolution operation, I removed the restriction of fixed sizing. So now we are able to feed the entire image as input to the convolution neural network and combined with the concept of the image pyramid, I can get the spatial output that is by using the images of different sizes in different scales. I'll be able to detect different objects of different sizes at different locations so both my problem the problem of location and scale are solved.

Final Deliverable of the Project

HW/SW integrated system

Core Industry

Transportation

Other Industries

Others

Core Technology

Artificial Intelligence(AI)

Other Technologies

Augmented & Virtual Reality, Others

Sustainable Development Goals

Decent Work and Economic Growth, Sustainable Cities and Communities

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com