The era of autonomous vehicles is rapidly approaching, bringing with it the imperative need for advanced detection technology to guarantee the safety of pedestrians and drivers. Daniel Gehrig and Davide Scaramuzza, esteemed researchers from the Department of Informatics at the University of Zurich (UZH), have devised an innovative solution by integrating event camera technology with artificial intelligence (AI) to enhance the effectiveness of pedestrian detection systems.
The existing pedestrian detection systems in vehicles have exhibited shortcomings, as evidenced by the alarming number of vehicular fatalities in the United States – the highest in 41 years, reaching 7,508 in 2022. Conventional frame-based systems, capturing 30-50 frames per second, while capable of visually or audibly alerting the driver, may prove inadequate in identifying imminent obstacles between frames.
Addressing this challenge, Gehrig, the lead author of the published paper in Nature, underscored the complexities of increasing frame rates, necessitating real-time processing of additional data and heightened computational power. This is where event cameras, a recent innovation, prove indispensable. Distinguished from traditional cameras, event cameras feature intelligent pixels that record information upon motion detection, eluding gaps between frames and enabling swift obstacle detection. Scaramuzza noted the resemblance of event cameras to human eyes in image perception, hence their alternate moniker – neuromorphic cameras.
Though event cameras present limitations, particularly in perceiving slow-moving objects and converting data to train AI algorithms, Gehrig and Scaramuzza devised a strategy to amalgamate the strengths of both systems. Their hybrid solution incorporates a standard camera capturing 20 images per second and an AI system known as a convolutional neural network to recognise objects. The event camera data is subsequently processed by an asynchronous graph neural network renowned for its proficiency in analysing 3D evolving data.
The outcome is a visual detector matching the swiftness of a standard camera capturing 5,000 images per second, with the same bandwidth requirement as a standard 50-frame-per-second camera. Furthermore, the system adeptly identifies vehicles and pedestrians entering the field of view between subsequent frames of the standard camera, thereby enhancing safety for all, particularly at high speeds.
The future appears increasingly promising with the scholars contemplating the integration of LiDAR sensors into their design, a feature presently utilised in self-driving vehicles. This advancement holds the potential to elevate the system’s performance to unprecedented levels.
In essence, the collaborative endeavours of Gehrig and Scaramuzza have yielded a pioneering solution poised to redefine pedestrian detection in car sensors. As we approach the era of autonomous vehicles, their contribution will undoubtedly prove pivotal in guaranteeing the safety of pedestrians and drivers on the roads.