Engineers at Yunnan University in China have developed a LIDAR-based automatic object recognition and tracking system for smart vehicles that can identify cars, objects, and pedestrians. The idea is to enable cars to sense their surroundings while simultaneously communicating with ITSs (intelligent transportation systems). This would provide useful functions like automatic obstacle detection, dynamic cruise control, and self-driving. Moreover, these capabilities can help prevent accidents and reduce traffic congestion substantially.
Today’s smart vehicles struggle to accurately make sense of their surroundings for various reasons. For example, mounted cameras often have a limited range and perform poorly in adverse conditions, such as low light, fog or rain. Additionally, identifying vehicles, pedestrians, and other obstacles in images is computationally intensive, requiring a lot of time and processing power. What’s more, most image processing algorithms tested in smart cars can misclassify pedestrians, other vehicles, and static objects such as trees.
The researchers looked to mitigate those issues by replacing those mounted cameras used to identify and track its surroundings with LIDAR. “Put simply, the working principle of LIDAR involves irradiating a target with a laser and detecting the radiation that is echoed back through the processes of reflection, refraction, scattering, and transmission,” explained graduate student Xiangtian Zheng. “Most importantly, by measuring the time it takes for the laser to return to the emitter after encountering an obstacle, we can accurately calculate the distance between the object and the device within a resolution of mere centimeters.”
Beyond greater range and accuracy, LIDAR data can be easier to process over sets of images. To that end, the team used grid maps with the idea of first making a spatial map of all the distances measured by the LIDAR. The premise is that consecutive points on this map, whose difference in distance is small, most likely correspond to the same object. This allows an algorithm to divide the map into regions that contain each individual object in the field of view. By removing other features from the LIDAR data and combining certain filtering and classification techniques, the model can tell vehicles apart from pedestrians and static objects. It can also estimate the trajectories of the objects it identifies, which is a crucial feature for autonomous driving.
The system was tested against a conventional Lidar-based platform and gauged the capacity of both systems to accurately identify multiple objects at different distances, reliably detect pedestrians, and track multiple moving vehicles simultaneously. The engineers found that their system produced higher recognition rates for vehicles and pedestrians and increased accuracy in estimating tracking trajectories. It will be interesting to see how this new approach handles real-world tests in the near future.