Lean Mean Object Detection Machine

In recent years, autonomous navigation has seen a remarkable surge in innovation, enabling many technological advancements in automobiles, drones, and other robotic systems. A key factor in the success of these autonomous navigation systems is the use of sophisticated object detection models. These models play a vital role in enabling machines to perceive and comprehend their surroundings, allowing for safe and efficient navigation in complex environments.

Object detection is vital to autonomous navigation, as it allows machines to identify and classify various obstacles, pedestrians, vehicles, and other relevant entities in real-time. This capability is essential for making informed decisions and taking appropriate actions to navigate through dynamic and unpredictable scenarios. The ability to detect and react to a variety of objects in the environment is a key factor in ensuring the safety and reliability of autonomous systems.

One of the challenges in deploying object detection models, such as the popular You Only Look Once (YOLO) algorithm, lies in their need for substantial computational resources. These models often demand significant computing power, making them impractical for many applications due to issues of cost, bulkiness, and high energy consumption. As a result, there is a growing demand for more efficient and lightweight object detection models that can strike a balance between accuracy and resource efficiency, enabling widespread adoption across a range of autonomous systems.

Researchers at the University of Maryland and Johns Hopkins University recently teamed up to build a more efficient object detection model that could help to fill this present need. Called Squeezed Edge YOLO, their object detector was designed to run on tiny edge computing platforms. As the name implies, the model was squeezed down to a miniature size, in the kilobyte range, which has dramatically increased both inference speeds and energy efficiency when compared with traditional YOLO models that have been optimized for edge machine learning.

To achieve their feat, the researchers focused on optimizing their model for the GAP8 hardware architecture, which consists of a primary microcontroller, a secondary octacore processor, and a number of hardware accelerators, like a convolution engine. As a first step, they began with the EdgeYOLO model, and worked to shrink down the size of the input images so that they could fit within the memory of their GAP8-based development board. Further, the team reduced the number of input and output channels present in the convolutional layers and, where necessary, also reduced the size of the kernel. Finally, a number of residual blocks were either removed or simplified, as they would otherwise excessively tax the GAP8 hardware.

This novel algorithm was tested on a pair of edge computing platforms — an AI-deck with a GAP8 microcontroller and an NVIDIA Jetson Nano with 4 GB of RAM. The AI-deck powered a Crazyflie drone, while the Jetson was used as a controller for a JetBot. After training the Squeezed Edge YOLO model on over 8,000 images, its object detection capabilities were assessed. As compared to EdgeYOLO, the new system ran 3.3 times faster, and did so while consuming 76% less energy. Moreover, Squeezed Edge YOLO is eight times smaller than EdgeYOLO.

These advantages did not come at the expense of accuracy. The object detection capabilities of the new model were not significantly different from larger models. This combination of accuracy and efficiency could enable Squeezed Edge YOLO to be used in a wide range of autonomous vehicles in the future.