An interesting debate has been made between researchers and organizations about the role of LiDAR and computer vision in future autonomous cars.
LiDAR is a type of sonar used for principle light detection and ranging by finding the precise distance between its source and nearby objects. This is quite a helpful feature for self-driving cars to read and understand their environment. Its range lies within 60 meters and is highly ideal for 3D mapping.
LiDAR technology also shows potential improvement with the help of solid-state sensors, which increase the sensor range up to 200 meters. Moreover, 4-differential lidar can determine the position and velocity of objects in a 3-dimensional domain. One of the most significant current concerns about lidar is its high cost.
However, lidar is not the only option for autonomous vehicle development. Leading sustainable energy AV and EV developer Tesla preferred the implementation of cameras to improve self-driving vehicles.
Another concern facing AV design is safety. That’s where computer vision, a branch of machine learning widely used to lower the risk of driving obstacles for vehicles, comes in.
Although, this domain is still growing, and self-driving vehicles can’t be consistently tested in urban settings yet. For many developers in the AV space, computer vision is preferred for the safety of pedestrians and passengers.
Evaluating Lidar Sensing Capabilities
In a basic functional sense, lidar sends a laser signal out and waits until it is reflected back to its photodetectors, which measure the difference of time to determine the object’s distance from its sensors.
The greater the time it takes, the further the object is. Moreover, lidar uses infrared light, making it more vulnerable to environmental conditions such as fog or rain.
Lidar sensors have the advantage of not being dependent on ambient light; as they can generate their own pulses. Which helps them better detect small objects with high precision. However, lidar is often also unreliable at night or in inclement weather.
Overall, lidar sensors offer several advantages for autonomous vehicle development. One of the most effective approaches is 3D cloud points – used to work with reflective and both textured and non-textured surfaces.
In terms of cost, in 2012, Google introduced its AV prototype utilizing a lidar system valuing $70,000. Costs are significantly lower now, enabling smaller tech firms more opportunities to invest in the technology.
Making a Case for Computer Vision
There are many advantages of the lidar system, but on the other hand, cameras show a preferable trend towards price and text/color recognition.
Lidar works great for spatial information but currently has trouble differentiating between a serious obstacle in a vehicle’s path with something less concerning.
Can they have told us that a pedestrian is focusing on this phone right now? The answer to this question is no. In adverse environmental conditions, cameras with any simple radar may perform better, they’d also likely be more cost-effective.
In the future, computer vision technology must interpret data faster and analyze its environment comprehensively. For now, one of the best hybrid approaches is the dual use of lidar and computer vision to detect objects, color, and text to get a clear picture of any vehicle’s environment. Nevertheless, this method is totally dependent on the programmer, and it is also computationally expensive.
Currently, the progress trend is a race between artificial intelligence programmers and lidar cost curve. If the industry succeeded in creating such a neural network that would get reliable and quick image information, then lidar would become an additional feature.
In the case of deep learning models, there are still limitations for large training datasets and computation power. In order to achieve higher accuracy for DL models, training datasets must be enough to generate accurate results.
Scaling computer vision systems with PyTorch is one notable and advanced method to optimize deep learning models. PyTorch/XLA is a python library that facilitates accelerated linear algebra to connect with cloud TPUs.
Though AV development still faces some limitations, thanks to the work being put into lidar and computer vision processes, a future where self-driving technology is self-reliant, intelligent, and reliable is within reach.