ColorByte: A real time MOT method using fast appearance feature based on ByteTrack

Abstract

Multi-object tracking (MOT) aims at estimating bounding boxes and identities of objects in videos. With the wide application of embedded devices, MOT methods with acceptable accuracy that performed in real-time on these weak devices is becoming more and more important. In order to realize real-time MOT on embedded IoT devices, the best trade-off between accuracy and inference latency is the key to achieve. To achieve high accuracy, popular MOT methods introduce Re-ID module to integrate with the detection-based MOT method and train two DNNs simultaneously. However, the integration causes a conflict between computing cost and training both tasks to achieve good results. To address this key issue, we design a fast appearance feature, which is a simple but relatively accurate method, to substitute cumbersome Re-ID component. Besides, ByteTrack is the new SOTA association algorithm in MOT benchmarks which introduce an extra association on objects with low score. Based on ByteTrack, we propose an improved association method to remove most of the background interference based on the results from appearance extraction and recover part of lost detection boxes after the association based on IoU. In addition, we turn down the detection threshold and release more boxes for the low sensitivity of our own feature extraction method. We evaluate our methods and achieve 70.1 MOTA, 81.9 MOTP and 69.2 IDFI with real-time running speed on NVIDIA Xavier NX.