Abstract
This paper proposes a novel motion field estimation method based on a 3D light detection and ranging (LiDAR) sensor for motion sensing for intelligent driverless vehicles and active collision avoidance systems. Unlike multiple target tracking methods, which estimate the motion state of detected targets, such as cars and pedestrians, motion field estimation regards the whole scene as a motion field in which each little element has its own motion state. Compared to multiple target tracking, segmentation errors and data association errors have much less significance in motion field estimation, making it more accurate and robust. This paper presents an intact 3D LiDAR-based motion field estimation method, including pre-processing, a theoretical framework for the motion field estimation problem and practical solutions. The 3D LiDAR measurements are first projected to small-scale polar grids, and then, after data association and Kalman filtering, the motion state of every moving grid is estimated. To reduce computing time, a fast data association algorithm is proposed. Furthermore, considering the spatial correlation of motion among neighboring grids, a novel spatial-smoothing algorithm is also presented to optimize the motion field. The experimental results using several data sets captured in different cities indicate that the proposed motion field estimation is able to run in real-time and performs robustly and effectively.
Highlights
The accurate perception of the motion of moving objects is a key technology for active collision avoidance systems and intelligent driverless vehicle systems
In Multi-target tracking (MTT), each neighborhood system is regarded as a moving target, and its position and velocity are filtered by the linear Kalman filter, while, in MFE, for each moving neighborhood system, the average of the velocities of its constituent grids is calculated as its mean velocity, which will be compared with the velocity of the corresponding target in MTT
We propose a motion field estimation method based on a 3D light detection and ranging (LiDAR) and present both the theoretical framework and a practical solution
Summary
The accurate perception of the motion of moving objects is a key technology for active collision avoidance systems and intelligent driverless vehicle systems. Multi-target tracking (MTT) ([1,2,3,4,5]) is one effective method to accomplish this work, which estimates the motion state of all of the detected targets, such as vehicles, pedestrians, and so on. MTT is a well-theorized and intuitive motion sensing method It includes three main steps, that is object segmentation, data association and motion estimation. The remainder of this paper is organized as follows: Section 2 describes the pre-processing of raw measurements; Section 3 details the global motion field estimation algorithm, including data association, motion state estimation and spatial smoothing; Section 4 demonstrates the experimental results in real-world scenarios; Section 5 offers some conclusions and remarks
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
