Efficient Packaging Line Object Counting by Cross-Frame Association With Wavelet Convolutions and Trajectory Compensation

Longxuan Wei, Yutao Zhu, Yufeng Li, Ming Qian, Xiang Zuo, Boan Chen, Shiyu Liang, Zhouhan Lin, Junchi Yan

Abstract: Real-time object counting in the industry pipeline is critical for improving efficiency and accuracy in industries like manufacturing and logistics. This paper introduces a novel multi-object association method, namely tracking method, which is specifically designed for real-time object counting in industrial environments. We present a specialized object counting dataset and a joint method that enhances multi-object tracking performance by focusing on both detection and tracking. Our approach features WT-YOLO for object detection, a Wavelet Transform-based convolutional neural network that leverages the wavelet transform to capture both spatial and frequency domain information. WT-YOLO significantly improves detection accuracy in complex scenes with occlusions and irregular shapes by expanding the receptive field of the network and capturing low-level shape signals, thereby enhancing the network’s shape bias. Experimental results demonstrate that WT-YOLO significantly outperforms YOLO-v8 and YOLO-MS on both the object dataset and a damaged COCO dataset, even with similar parameter settings. Additionally, we introduce the Virtual Trajectory Compensation (VTC) module, which adapts to the non-linear motion of objects on packaging lines by enhancing the parameter updating of motion models during re-correlation. The VTC module is a plug-and-play addition that can be seamlessly integrated into Kalman filter-based trackers to mitigate tracking loss issues. Our method achieves high accuracy and speed, making it suitable for real-time applications. Our work offers new insights and advancements in multi-object tracking for industrial settings by addressing the challenges of occlusions, irregular shapes, and non-linear motion in object counting tasks.