Online Multiple Object Tracking Using Rule Distillated Siamese Random Forest

Abstract

In a multiple object tracking (MOT) system, an association check between the tracker and detected objects is an important factor in determining the tracking performance. Siamese convolution neural network (CNN) is the most popular data association method in MOT owing to its good matching performance and network sharing support. However, it is unsuitable for real-time online tracking in low-end systems because numerous parameters and operations are still required. In this paper, instead of a CNN, we propose using a SiameseRF algorithm which combines Siamese structure and random forest (RF), enabling highspeed learning and classification. SiameseRF has a shared-rule based Siamese structure rather than shared weight, which improves the matching performance and solves existing slow CNN-based tracking issues. During the learning process, the shared RFs consisting of tree rules are learned in the directions of increasing similarity to the positive pair {anchor, positive} and increasing difference between the negative pair {anchor, negative}. However, because many rules that make up SiameseRF remain a burden for online processing, this study proposes an additional rule distillation algorithm to effectively remove redundant and unimportant rules causing an overfitting with SiameseRF. This reduction in the number of rules reduces the processing time and number of parameters in the rule distilled SiameseRF. In experiments conducted on MOT benchmark datasets, our proposed 30% rule distilled SiameseRF achieved up to a 1.12-times faster speed and a 1.13-times higher compression rate than basic SiameseRF while maintaining a similar or somewhat better tracking performance than other state-of-the-art CNN-based MOT algorithms.

Graph Convolution Neural Network-Based Data Association for Online Multi-Object Tracking

“…Inspired by [10], we defined the online MOT matching rule as follows: When the object of detection does not match any tracker, this detection is assigned as a new potential tracker, and if the potential tracker matches more than τ times, it is assigned as an actual tracker. Otherwise, it is declared a false tracker and removed.…”

“…The key issue in real-time data association is determining the optimal association between detections and trackers. The most representative data association methods are bipartite assignment [5,6] and Hungarian-based approaches [7][8][9][10]. These methods model weights as an affinity matrix between graph sets consisting of existing trajectory nodes and new detection nodes [11].…”

“…Deep learning technologies applied in a data association for MOT have recently achieved a state-of-the-art performance. The Siamese network [10,[21][22] and its modified triplet [23][24] and quadruplet network [25] applied the same network to the detection and tracker and calculated the similarity based on the difference in the output characteristics. The more complex the structure of the shared network is, the higher the matching performance; however, real-time online tracking becomes more difficult.…”

Real-Time Multiobject Tracking Based on Multiway Concurrency

“…The average frame rate can reach 10 fps. In [19], the author combined Siamese structure and random forest (RF) and improved the matching performance and solved existing slow CNN-based tracking issues for MOT via a shared-rule based Siamese structure. The algorithm considers both performance and real-time, the processing speed was 12.4 fps.…”

Appearance Guidance Attention for Multi-Object Tracking

“…Appearance feature vectors are usually learned by neural networks and guided training through loss functions. At present, many models use Siamese network [22] [27] or recurrent neural network for appearance feature learning and object association [17] [19]. However, these methods will introduce additional convolution neural networks in the tracking paradigm, which increase computation cost and impair tracking speed.…”