Overcoming Perception and Depth Limitations for Safer Roads and Autonomy

The Difficult Reality of Road Safety

Improved ADAS and AV deployment will reduce road traffic injuries. Accidents continue to be a leading cause of death and injuries worldwide, with over 1.3M deaths and between 20-50M injuries estimated annually, and are by far the leading cause of preventable injury ([1] WHO, 2022, [2] NSC, 2022). They are the top cause of death for young people aged 5-29 years and more than half of all road traffic deaths are among vulnerable road users: pedestrians, cyclists, and motorcyclists. This sobering statistic remains high, accompanied by an estimated economic impact of road accidents at 1-2% of a country’s GDP annually.

In the US alone, traffic deaths have reached a 16-year high and pedestrian fatalities reached their highest level in 40 Years (GHSA, 2022), with nighttime pedestrian deaths rising by 41% since 2014. Specifically for trucking, there were nearly 550,000 accidents annually in the US, with approximately 130,000 causing injuries, 100,000 caused by sleep deprivation, and a 54% increase in fatalities involving trucks since 2009.

Growing Truck Driver Shortages

One clear opportunity is commercial transportation. Much has been written about the worldwide truck driver shortage. A new International Road Transport Union study (IRU, 2022) shows an estimated 2.6M unfulfilled driver positions across the 25 countries surveyed, with shortages of 80,000 in the US, nearly 400,000 across Europe, and 1.8M in China. A significant increase is forecast for 2022 and expected to grow due to a lack of skilled workers, an aging driver population, and waning interest in the profession due to quality of life issues. Coach and bus driver populations showed similar trends.

Robust Perception and Depth is Fundamental for ADAS and AVs

The perception stack is the most crucial element of the AV or ADAS system for improving vehicle safety. It ingests sensor data and interprets its surroundings through an ensemble of sophisticated algorithms [Figure 2]. It then delivers the information needed for the vehicle to warn a driver or make good control decisions itself.

For an L2 – L3 (SAE, 2021) driving system to help a driver, or an L4 autonomous vehicle safely drive itself, the perception system must provide robust detection and accurate distance estimation of objects in all lighting and weather conditions (e.g. rain, fog, or snow) and in city or highway routes. It can be constrained to the operational design domain (ODD) where the system was designed to properly operate, such as only during highway travel or only under well-lit daytime conditions. Interestingly enough, the automotive industry calls low light and poor weather “corner cases” despite being common occurrences. It must have sufficient range to provide time for the system or driver to respond to an issue. At highway speeds, this needs to be out to 300m for passenger vehicles and over 500m for Class 8 trucks in order to safely brake or maneuver to avoid an accident.

ADAS sensing today primarily relies on cameras and radar to enable detection and distance estimation for driving functions. Sensor data is processed with AI and classical perception algorithms to detect pedestrians, vehicles, lanes, and signs, and determine position, obstacle range, and velocity. This information is used to deliver driver warning or initiate active control of the vehicle, such as automatic braking to avoid a collision.

AVs have the much more complex objective of enabling driverless operation across the vehicle’s ODD. The vehicle is now responsible for correct and safe operation rather than a driver. As such, companies have heavily invested in vehicle architectures with extensive and costly sensor suites that include many cameras, radars, numerous lidar, and AI software autonomy stacks that handle the full operation of the vehicle. Recently, AV leaders have begun deploying vehicles for driverless paid rides, long-haul freight, and last mile deliveries on the path toward commercial viability.

But there is still a road ahead for these companies to cost-, power-, and performance-optimize the hardware and software to achieve scalable effective deployment.

The unfortunate reality is that many of today’s production ADAS systems fail under the “corner case” scenarios and even struggle in good conditions, as highlighted again by the AAA (AAA, 2022). While much more capable than their ADAS counterparts, AV platforms are also hampered by the same challenges and in reality are still in the prototype and testing stage.

This accuracy and robustness gap continues to be confirmed in benchmarks during Algolux’s OEM and Tier 1 customer engagements. These ADAS / AV systems have a reduced ability to properly detect objects and estimate depth in driving scenarios where there is high sensor noise, low contrast, occlusions, and dealing with previously unseen objects and orientations.

Addressing Key Perception and Depth Gaps

Cameras are the predominant external sensor for ADAS and these L2 – L3 systems are expected to use up to 20 cameras (Yole Developpement, 2022) from 1.3MP to 8MP. They are also an increasingly important component in an AV system’s sensor suite [Figure 2].