Algorithms for Autonomous Vehicles Take Decisions Based on Data

Knowledge Hub
Automotive Performance Engineering

Powered by

Algorithms for Autonomous Vehicles Take Decisions Based on Data


There is a great deal of buzz generated over autonomous vehicles or self-driving cars across the automotive industry globally. However, in the Indian context, there are reservations over adoption of autonomous vehicles or self-driving cars. The general line of thought is that autonomous vehicles or self-driving cars would end up eliminating jobs of drivers in a large way and India will not be ready with required infrastructure in years to come. Sharing his perspective on adoption of autonomous vehicles, Prashant Deshpande, Founder & Managing Director, EC-Mobility (a company that offers future mobility solutions in Advanced Driver-Assistance Systems (ADAS) and Autonomous Driving among others), said autonomous vehicles are no longer considered a futuristic dream and could turn into reality sooner than many among us probably think.

Given the transformations witnessed across the automotive industry, the EC-Mobility’s top official said going forward the industry will be hugely dictated by usage of software and data. Automotive OEMs and Tier 1 suppliers across the globe are working towards achieving Level Four and Level Five automation but how these levels of automation pan out remains to be seen, he noted.

The rapidly-evolving automotive industry will see data emerge as the new ‘oil’ and the onus will be on automotive companies to turn these data into value, Deshpande said. Vehicles in contemporary times need oil to drive but tomorrow it will need data to move.

Autonomous cars typically necessitate the need to perceive a situation or scenario and accordingly plan and act. Deshpande said the most important step is to interpret with precision if the user is launching any driver assistance feature, as it is critical to ensure data interpretation, as distinguishing between perception and interpretation holds the key. For algorithms to understand there is a need to carry out labelling and even that is not enough because algorithms has to take decisions based on data, Deshpande explained.

It may be noted that tonnes of data are being generated in the automotive space. For autonomous driving or any driver assistance feature, every care must be taken to ensure data is correctly labelled and prepared for a vehicle’s algorithm to take the right driving decisions, the EC-Mobility MD observed.

In addition to box labelling, 2D semantic segmentation is another key focus area of autonomous vehicles. Deshpande said the vehicle’s algorithm will not be able to define a scenario in isolation unless semantic segmentation of that particular scenario is carried out.

He also threw light on semantic labelling for driver behaviour monitoring. “If a driver is sleepy or not adhering to traffic rules or driving instructions, one needs to look at his eye movements also. Eye tracking can be captured in a video. Interpreted data is valuable data and without data labelling one cannot make interpretations and without interpretations one cannot arrive at logical decisions,” Deshpande quipped.

Clearly, correct interpretations are critical and that explains why so much of data is captured in various scenarios such as day/night, rainy/snowy conditions and from various sensors like RADAR, LIDAR, Cameras, etc. Deshpande said all scenarios need to be captured in order to provide intelligence to data that will enable users to run algorithms for driver assistance features like lane departure warning, parking assistance, collision detection warning, etc. The machine or algorithm has to learn and such learning happens when learning data is provided to the machine or algorithm. Deshpande said it’s all about offering inputs to the vehicle’s algorithm and the algorithm will come to the conclusion that the pedestrian will cross the road or not based on the data that is available.

Deshpande said whether the vehicle’s sensor is performing correctly or algorithm is performing correctly can be determined by comparing trained and untrained data. In case the vehicle’s algorithm is not performing correctly and data is not labelled, a vehicle user will encounter driving challenges, he added.

Deshpande has no doubts that data is going to be crucial towards make L3, L4 and L5 a reality and this data has to be enriched every time one has new sensors, new cars and new situations. The need for refined data is equally important for autonomous cars. Deshpande said in scenarios where the vehicle has sensor data, there is a need to carry out annotation on the ground truth data and that is the value addition required. A process called ground truth labelling is used to annotate recorded sensor data with the expected state of the automated driving system. The test vehicle needs specialised software tools for labelling, annotation and for various other functionalities, he noted.

Deshpande said it is critical for autonomous cars to ensure control and execution as there is no way one can work in isolation, as one must have a perception where the data will be used and how critical is that aspect. There are tools available where in a guided environment one first does it manually, then perform semi-automation and then achieve automation, in terms of data annotation or data preparation, he noted.