Machine Learning is essential for the success of safety-critical, holistic perception. At the same time, learned approaches need investment to achieve flexibility.
In the ever-evolving landscape of 3D sensor perception, this statement encapsulates the critical balance between traditional rule-based algorithms and deep learning approaches. While the automotive industry has been the cradle of safety-critical 3D sensor perception, the static sensor sector has leaned on traditional methods.
Static sensor sector includes anything and everything where the 3D sensors are installed on the infrastructure rather than a moving vehicle with the main applications being: people management, intelligent transport systems, smart roadside infrastructure, security and access control. This choice of relying on traditional perception algorithms in the static sensor sector is pragmatic given lower volumes per opportunity. However, such traditional approaches are over-reductionist and result in compromised accuracy as they fail to harness the true power of 3D sensors. Deep learning, on the other hand, promises high-fidelity perception yet it is not primarily in vogue because it comes with its own set of challenges. Thus, the future of 3D sensor perception lies in harnessing the high precision and accuracies of learned approaches for 3D sensor perception along with reduction of the associated challenges of data and deployments.
The Legacy of Traditional Rule-Based Algorithms
Traditional rule-based clustering and tracking algorithms were initially conceived in the automotive industry. They always essentially serve as safety nets, ensuring to take obvious decisions like slamming of brakes if there is an object right in front of the car, no matter what the object is. Because these algorithms do not have the overhead of data and engineering, they have held their ground in non-automotive, static sensor applications. Their limitations however become apparent as demands for high accuracy and precision increase. Hence, POCs and lab successes fail to translate into large-scale deployments, leaving a gap in the market for a more robust solution.
High fidelity of Learned Approaches
Deep learning emerged as a game-changer by learning the rules to detect objects from the complexities of the real-world data. It offered the high-fidelity detections required for a wide array of applications, whether safety-critical or not. However, this approach too is not without challenges of its own. Deep learning networks demand vast amounts of annotated or labelled data as fuel to operate effectively. Yet, 3D sensor technology remains niche, resulting in a scarcity of research and open-source state-of-the-art networks and large-scale datasets. Annotating 3D sensor data is a laborious and costly process, and deep learning networks lack the transparency required for safety-critical applications.
The Cron AI Approach
At Cron AI, we firmly believe that the future of 3D sensor perception lies in the synergy of learned and traditional approaches, particularly for safety-critical applications. While traditional algorithms provide the necessary guarantees and safety nets, high-fidelity detections for vehicle navigation and steering are the domain of learned approaches. Safety nets are necessary when the system behaviour moves into an unknown zone, for example, unknown objects or objects which are large and very close to the vehicle. The obvious behaviour in such cases is to slam the brakes and stop the car. Since these safety nets are critical to mitigate accidents and not to navigate or steer the vehicle, their ownership and onus always lies with the Car OEMs and Tier 1s. Designing of such safety net traditional algorithms is not our focus at Cron AI.
Our innovation is guided by the mission to make deep learning-based 3D sensor perception more flexible, cost-effective, and quicker to deploy, all while enhancing interpretability. For non-automated driving applications, i.e. the static sensor industry, traditional algorithms completely fall off the cliff. Even though these applications are not safety critical, they are always more complicated than an automated driving application, with the caveat that a failure will not always lead to loss of lives. All intelligent transport applications, smart roadside infrastructure applications, people management applications in travel hubs, airports, smart spaces are high object density applications where objects are occluded, partially visible, do not reflect back the LiDAR IR rays, noisy and full of distortions. Such high complication scenarios require a high fidelity and accurate 3D perception system which can harness the true power of 3D sensors.
The Power of Cron AI’s Innovative Deep Neural Network
We’ve developed a patented deep neural network based 3D perception system driven by self-supervision, simulations, and supervised learning at Cron AI. This approach maximises the DNN learning capacity, enabling our network to glean more insights from less data. For our network it’s not just about recognising shapes of the objects; it’s about comprehending the context of the 3D sensor image, much like how humans perceive the world. Our network does not ‘just’ see a car as an object with wheels and a chassis; it understands the entire scene, recognising the road, the cars, the sidewalk, the trees, and the pedestrians. This inherent understanding of point clouds and their context makes our 3D perception robust in real-world scenarios. At a very high level, our network has the capability to understand what is actually going on in the real world when it looks at a LiDAR image rather than just looking for objects which have the shape of cars or humans as it has learned before.
Efficiency Meets Precision:
Our deep learning network is one of the first universal yet state of the art network in 3D sensor perception industry. It is optimised for high accuracy, precision, and recall while maintaining efficiency and fast execution. This ensures that our solution is not a mere theoretical marvel but is practical, capable of being deployed on the field in accordance with customer requirements. It’s not just about high-fidelity detection; it’s about doing so efficiently and in real-time, making our solution suitable for a wide range of applications.
Conclusion
The future of 3D sensor perception is not a choice between traditional rule-based algorithms or deep learning. Safety-critical applications require the guarantees and safety nets provided by traditional algorithms, while the high-fidelity detections that deep learning offers are essential for vehicle steering and navigation. At Cron AI, we’ve embarked on a journey to create a holistic deep learning based 3D sensor perception. By democratising learned approaches, we’re not only advancing 3D sensor perception but paving the way for a safer and more efficient world.
About the author: Saurav Agarwala is the Co founder & CTO at Cron AI. He is a deep learning and lidar perception enthusiast.
