With recent advances in object detection, the tracking-by-detection method has become mainstream for multi-object tracking in computer vision. The tracking-by-detection scheme necessarily has to resolve a problem of data association between existing tracks and newly received detections at each frame. In this paper, we propose a new deep neural network (DNN) architecture that can solve the data association problem with a variable number of both tracks and detections including false positives. The proposed network consists of two parts: encoder and decoder. The encoder is the fully connected network with several layers that take bounding boxes of both detection and track-history as inputs. The outputs of the encoder are sequentially fed into the decoder which is composed of the bi-directional Long Short-Term Memory (LSTM) networks with a projection layer. The final output of the proposed network is an association matrix that reflects matching scores between tracks and detections. To train the network, we generate training samples using the annotation of Stanford Drone Dataset (SDD). The experiment results show that the proposed network achieves considerably high recall and precision rate as the binary classifier for the assignment tasks. We apply our network to track multiple objects on real-world datasets and evaluate the tracking performance. The performance of our tracker outperforms previous works based on DNN and comparable to other state-of-the-art methods.

In this paper, we propose an efficient online multi-object tracking method based on the Gaussian mixture probability hypothesis density (GMPHD) filter and occlusion group management scheme where a hierarchical data association is utilized for the GMPHD filter to reduce the false negatives caused by missed detection. The hierarchical data association consisting of two modules, detection-to-track and track-to-track associations, can recover the lost tracks and their switched IDs. In addition, the proposed grouping management scheme handles occlusion problems with two main parts. The first part, "track merging" can merge the false positive tracks caused by false positive detections from occlusions. The occlusion of the false positive tracks is usually measured with some metric. In this research, we define the occlusion measure between visual objects, as sum-of-intersection-over-each-area (SIOA) instead of the commonly used intersection-over-union (IOU). The second part, "occlusion group energy minimization (OGEM)" prevents the occluded true positive tracks from false "track merging". Each group of the occluded objects is expressed with an energy function and an optimal hypothesis will be obtained by minimizing the energy. We evaluate the proposed tracker in benchmarks such as MOT15 and MOT17 which are public datasets for multi-person tracking. An ablation study in training dataset reveals not only that "track merging" and "OGEM" complement each other, but also that the proposed tracking method shows more robust performance and less sensitiveness than baseline methods. Also, the tracking performance with SIOA is better than that with IOU for various sizes of false positives. Experimental results show that the proposed tracker efficiently handles occlusion situations and achieves competitive performance compared to the state-of-the-art methods. In fact, our method shows the best multi-object tracking accuracy among the online and real-time executable methods.

Tracking multiple objects in a video sequence can be accomplished by identifying the objects appearing in the sequence and distinguishing between them. Therefore, many recent multi-object tracking (MOT) methods have utilized re-identification and distance metric learning to distinguish between objects by computing the similarity/dissimilarity scores. However, it is difficult to generalize such approaches for arbitrary video sequences, because some important information, such as the number of objects (classes) in a video, is not known in advance. Therefore, in this study, we applied a one-shot learning framework to the MOT problem. Our algorithm tracks objects by classifying newly observed objects into existing tracks, irrespective of the number of objects appearing in a video frame. The proposed method, called OneShotDA, exploits the one-shot learning framework based on an attention mechanism. Our neural network learns to classify unseen data samples using labels from a support set. Once the network has been trained, it predicts correct labels for newly received detection results based on the set of existing tracks. To analyze the effectiveness of our method, it was tested on the MOTchallenge benchmark datasets (MOT16 and MOT17 datasets). The results reveal that the performance of the proposed method was comparable with those of current state-of-the-art methods. In particular, it is noteworthy that the proposed method ranked first among the online trackers on the MOT17 benchmark.