Video Object Counting With Scene-Aware Multi-Object Tracking

Introduction

With the development of object detection algorithms (Lin et al., 2017; Ayoun et al., 2010; Li et al., 2017; Zheng et al., 2019) and the innovation of reidentification technology (Wu, Wu et al., 2018; Lu et al., 2019; Zhang et al., 2019; Wu et al. 2018; Farenzena et al., 2010; Choe et al. 2019), tracking-by-detection has become one of the mainstream approaches for video analysis problems. It has been widely studied and has produced remarkable results in recent decades, demonstrating significant application value in vehicle navigation, intelligent monitoring, human-computer interaction, crowd counting, and other fields. We design a video object counting algorithm based on multi-object tracking based on the unique correspondence between the object and the ID number in the tracking algorithm. However, due to the presence of many uncertain factors in real-world scenes, such as occlusion, lighting change, scale change, and camera motion, maintaining the algorithm's accuracy and robustness remains a difficult task.

Figure 1.

The necessity of scene awareness. (1) Occlusion and deformation result in the complete change of the object’s appearance information and the loss of the original identity, which eventually leads to the statistical change of the number of objects. (2) We use the scene-aware method to determine the region where no object enters or leaves. Limit the increase in redundant IDs due to feature changes in this area to avoid an object being counted twice