Towards Latency-Stable Robotic Perception Pipelines in ROS 2: Eliminating Non-Maximum Suppression for Real-Time Target Tracking and Distance Estimation

Abstract

Visual perception is a fundamental component of autonomous
navigation in mobile robotics. However, for real-time control loops such
as Proportional-Integral-Derivative (PID) or Model Predictive Control
(MPC) systems, temporal determinism is as critical as spatial accuracy.
Traditional object detectors rely on Non-Maximum Suppression (NMS),
a post-processing stage whose execution time depends on scene complex-
ity. This dependency introduces stochastic inference latency, commonly
referred to as latency jitter. Such temporal instability propagates through
the perception pipeline, degrading multi-object tracking consistency and
significantly impairing monocular distance estimation, ultimately leading
to unstable robotic control behavior.
This paper proposes a latency-stable perception pipeline implemented in
Robot Operating System 2 (ROS 2) that fundamentally eliminates the
NMS bottleneck. By integrating an NMS-free detector (YOLOv10)
with the ByteTrack tracking algorithm inside an asynchronous multi-
threaded ROS 2 node, computational workloads are isolated from com-
munication callbacks, preventing head-of-line blocking and artificial la-
tency variance.
Experimental validation conducted in a dynamic Gazebo simulation envi-
ronment demonstrates a 78.2% reduction in latency standard deviation
(from 18.22 ms to 3.97 ms) compared to a classical baseline pipeline
(YOLOv8 + SORT). The improved temporal stability directly
enhances geometric consistency: the root mean square error (RMSE) of
monocular distance estimation decreases from 1.710 m to 1.449 m, while
the proposed system maintains errors below 1.5 m in 61.3% of frames
versus 26.1% for the baseline. The results confirm that eliminating NMS
is a key requirement for reliable deployment of vision-based perception
systems in real-time robotic platforms.