Moving Objects Tracking using Motion Vectors with Implementation on a Real ADAS Platform

The basic tasks of Advanced Driver-Assistance Systems (ADASs) are object detection and tracking. Although these tasks can be performed by processing signals from various in-vehicle sensors, the most used is the image captured by the in-vehicle camera. For an autonomous vehicle to know which action needs to be taken at a given moment, very useful data is about the speed and direction of movement of objects in their surroundings. Once objects are detected, different methods can be used to track them. Although there are various methods for object tracking based on computer vision and machine learning, they are very often not suitable for implementation on the real embedded ADAS platforms that have limited memory and computing resources. Therefore, this paper deals with the motion vectors (MVs) estimation between adjacent frames and their grouping with the goal of identifying and tracking objects around the ego-vehicle, all accompanied by the implementation of the proposed solutions onto a real embedded ADAS platform. Different methods for MV estimation and grouping are examined. Special attention was given to the efficient implementation of each task on the ADAS platform. The proposed solution achieves high performance in terms of accuracy when tested on real-life traffic videos. It is important to note that the solution is able to process 64 frames per second for the input images with a resolution of 1280x720 pixels.