Confidence Trigger Detection: Accelerating Real-time Tracking-by-detection Systems

TL;DR

This paper introduces Confidence-Triggered Detection (CTD), a novel approach that strategically skips object detection for frames exhibiting high similarity, leveraging tracker confidence scores, and identifies an optimal trade-off between tracking speed and accuracy.

Abstract

Real-time object tracking necessitates a delicate balance between speed and accuracy, a challenge exacerbated by the computational demands of deep learning methods. In this paper, we propose Confidence-Triggered Detection (CTD), an innovative approach that strategically bypasses object detection for frames closely resembling intermediate states, leveraging tracker confidence scores. CTD not only enhances tracking speed but also preserves accuracy, surpassing existing tracking algorithms. Through extensive evaluation across various tracker confidence thresholds, we identify an optimal trade-off between tracking speed and accuracy, providing crucial insights for parameter fine-tuning and enhancing CTD's practicality in real-world scenarios. Our experiments across diverse detection models underscore the robustness and versatility of the CTD framework, demonstrating its potential to enable real-time tracking in resource-constrained environments.

Figures (2)

• Figure 1: illustrates the tracking outcomes for frames 55 to 61 utilizing the CTD approach. The analysis is illustrated with cropped image frames, where white bounding boxes denote predictions made by the Kalman filter, and blue bounding boxes indicate detections from the object detector. Between frames 56 and 60, no new detections were recorded, because the confidence score, derived from the comparison of the white and blue bounding boxes, exceeded the predetermined threshold. However, at frame 61, a new detection was prompted due to the confidence score falling below the threshold. This indicates a notable deviation between the prediction from the Kalman filter and the object detection results from the previous frame.
• Figure 2: illustrates the trade-off between speed (FPS) and accuracy (MOTA) using different confidence thresholds.100% Threshold: Require detection for each frame, considered as baseline. 0% Threshold: Same as fixed frame skipping, never triggers detection based on confidence score but triggers when reaching the maximum skip frame threshold. It highlights that CTD consistently achieves superior accuracy levels at the same speed compared to the fixed frame skipping strategy (0% confidence threshold).