Sasha Andrieiev is CEO & Founder at Jelvix | Digital Leader | Innovation Expert.

“The future of electric vehicles is golf carts, not Tesla,” declared the Harvard Business Review in 2015. And it looks like they weren’t wrong. While major automakers are spending time testing autonomous car concepts, promising their appearance “soon” and “next year,” electric golf cart modifications could be a faster way to achieve the goal of putting self-driving cars on roadways.

Golf carts are now used for more than just playing golf. They are currently the most popular mode of transportation for short distances and are more eco-friendly than any alternative.

According to Allied Market Research, the global golf cart market will hit $1.79 billion by 2028, growing at a 3.9% CAGR between 2021 and 2028.

Faced with the harsh reality that a future of fully autonomous cars is further away than promised, automakers and tech companies are pivoting to alternate uses for self-driving technology. Some turn to delivery robots, while others are helping deploy small, low-speed golf-cart-type machines for sites, farms or airports.

For example, Honda is testing a new golf-cart-type model. That so-called “micromobility device” is part of Honda’s efforts to help people who can’t drive themselves, like the elderly or Generation Z. In their current state, Honda vehicles are more like golf carts or UTVs. But Honda believes these machines will have their place in the future urban environment.

Autonomous and semi-autonomous cars are equipped with sophisticated systems for detecting and recognizing vehicles and objects, road signs, traffic lights and road lanes they encounter on the road. These systems require a sufficient amount of high-quality data that are annotated properly. Annotation is adding digital labels to images or videos using bounding boxes and defining other attributes to help ML models recognize objects detected by vehicle sensors and cameras. These digitally tagged images are used to “train” driverless computer systems to identify key features when displaying new and unlabeled data.

A real-time high-precision perception system manages the vehicle’s actions by extracting visual information from the images captured by the cameras. The perception system understands the scene and then provides the decision system with data, including the location of the obstacles, judgment of whether the road is travelable, the position of the lanes, etc.

This is the process of determining which class a road sign belongs to. For this, a public data set on Kaggle is used, containing over 50,000 images of various traffic signs divided into 43 different classes. It is suitable at the initial stage, and then companies must collect and mark up their data independently.

Technology can help drivers identify the status of traffic lights and quickly decide according to the recognized status. Currently, these methods are based on traditional algorithms using image processing and machine learning.

The algorithm, for instance, first extracts the red, green and yellow objects from the image using a straightforward traffic light detection system. Traffic signal symbols are used to confirm an object’s authenticity before different traffic signal types are assessed in order to reduce environmental interference.

Road boundary or line detection is among the most crucial but challenging tasks in self-driving car development. It includes the road’s localization, the determination of the relative position between the car and the road, and the analysis of the automobile’s heading direction.

Existing methods of lane detection are based on deep learning techniques. Compared to traditional algorithms that rely heavily on visual cues in certain environments, deep learning improves the network’s ability to perceive a scene by constantly optimizing neural network parameters, resulting in higher reliability and applicability.

This is another component of a computer vision project required for autonomous cars. This solution is not included in the package needed for autonomous golf cars but would be appropriate within the smart city concept.

Working on annotating autonomous vehicles seems interesting, but it has many pitfalls. Companies that decide to take on the project independently will spend most of their time on the theoretical part and correcting their own mistakes. Besides, if the laborers are not data scientists, creating clear guidelines on what should be labeled is quite difficult.

Beyond experience, professional annotation companies must have innovative tools to realize perfect pixel accuracy. While the technology for simpler annotations is available for amateur use, the technology required to accomplish annotations on radar and lidar data is limited and requires high expertise.

That’s why it’s worth outsourcing providers with data annotation specialists. It’s a way to join an autonomous market without troubling data annotation.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?