Computer vision is the science and technology that allows machines to gain vision/sight.
As a scientific discipline, computer vision is concerned with the theory and technology for building artificial systems that obtain information from images or multi-dimensional data.
The Computer Vision team in Sally Robotics primarily works on the development of Vision systems that act as eyes for the Autonomous vehicle. The work done in Computer Vision essentially deals with the images obtained from cameras mounted on the vehicle (be it a normal camera, or a 3D camera such as a stereo camera).
The focus of Sally Robotics is development of intelligent autonomous systems that are robust enough to make them work in every environment (Indian roads, places with improper road marks and traffic signals; as well as in adverse conditions of haze, rain, low light, etc.).
In Sally Robotics, we use State-of-the-Art Deep Learning approaches and also develop our own algorithms that help the cars to obtain vision and also aid the Deep Learning community with new ideas.
Some of the problems that we try to solve are of Object Localization, Object Detection, Semantic Segmentation, Instance Segmentation, Human Pose Detection, Driver Attention, Dehazing, Deraining and Denoising
The insights from the CV system will be used in the decision-making system in the autonomous vehicle to control and navigate it.
In just a year within inception, the CV team has published a paper in the ‘3rd International Conference on Intelligent Autonomous Systems (ICoIAS)’ on the topic of Image Segmentation on unstructured environments (predominantly Indian roads) - Complete Scene Parsing for Autonomous Navigation in Unstructured Environments.
Last year, the team also started to make progress in the detection of traffic signals, starting from Speed Breakers. The team always strives to push the limits of what is achievable by a student team, both in terms of application and research.
Simultaneous localization and mapping (SLAM) is a hot issue in the field of unmanned systems. It is an essential task for the development of autonomous robots as well as vehicles.
SLAM works on the principles of detection and matching of feature points, selection and matching of keyframes and loop closure detection. It aims at a fully autonomous answer to the question “Where am I?” by providing an autonomously built map.
In Sally Robotics we use LIDAR sensors, Stereo Cameras, and RGB-D cameras as our sensors to capture the environment data around our Kobuki Turtlebot.
Our current focus is on visual-based SLAM methods so that we can better collaborate with Computer Vision tasks.
Our past projects include implementation of various SLAM methods including RTABMAP and ORB-SLAM as well as a kobuki follower which accomplishes dynamic obstacle avoidance.
Navigation aims at global path planning, that discovers the optimum route from one point to another, as well as the more challenging task of local path planning which deals with obstacle avoidance, behaviour planning and motion planning.
We have been using classical methods based on D*lite, time-elastic-band (TEB), rapidly exploring random tree (RRT) for motion planning and control. Some of these have also been tried on the Kobuki robot or ROS( robot operating system) simulators.
But with Sally Robotics’ long-term goal of attaining automation on Indian roads, we are now doing research on reinforcement learning-based approaches to tackle the numerous complexities of such an environment.
It involves a learning agent which figures out how its environment works, based on the reward received on performing various actions. These models first need to be trained and tested on a simulator such as CARLA before deployment on a real car in the real world.
With respect to the end goal, we are at a preliminary stage wherein we focus on solving some of the sub-problems in virtual environments and improvising on the same through our research and innovation.