Multi-object tracking (MOT) is a technique that aims to detect multiple objects in a single frame and maintain their identities in a given video sequence. Currently, the tracking-by-detection paradigm stands out as the most effective approach for MOT, encompassing object detection, person re-identification (Re-ID), and data association. In this paper, we propose enhancements to the Re-ID and data association modules, emphasizing the pivotal role of appearance information in achieving robust tracking quality. To capture more discriminative features, we introduce a network based on partial-level features. We also employ reparameterization to remove residual connection blocks from the backbone network without altering the original output to enhance tracking accuracy while optimizing computational efficiency. Moreover, the utilization of bounding boxes significantly influences tracking performance. We introduce an adaptive threshold mechanism and incorporate gating functions for data association, ensuring efficient use of each bounding box while minimizing trajectory fragmentation and missed detections. Experimental results illustrate that our proposed method attains high detection and tracking accuracy, particularly excelling in challenging video sequences. Our approach demonstrates comparable performance to state-of-the-art trackers on various benchmarks, including MOT17 and MOT20, where MOT20 presents frequent and severe interactions and occlusions.

In the standard bag‐of‐visual‐words model, the relationship between visual words and geometric structure information embedding in Voronoi cells is important for expressing the topology of the feature space. However, this information is usually ignored by recent works. To overcome it, the authors proposed a hybrid heterogeneous structure model (HHSM), where local hyperspheres and local structure subspaces are applied to simulate the intrinsic structure of the feature space. Firstly, the local hypersphere is formed by choosing some links between parts of visual words, with the use of a proposed decision strategy derived from k ‐dense neighbour algorithm. In order to capture the geometric structure information around the visual word, they then construct the local structure subspace with the transformed PCA principal vectors of the visual features within a Voronoi cell. Finally, this study introduces a novel feature encoding method based on the HHSM. Experiments are conducted on 15‐Scenes, Pascal VOC2007, Caltech101, Caltech256 and MIT Indoor 67 datasets, which include 4485, 9963, 9146, 30607 and 15620 images, respectively. The results demonstrate the effectiveness of the proposed method in improving the accuracy of the classification. In addition, the proposed method achieves comparable performance when combined with CNN local features.

In natural images, it remains challenging to estimate dissimilarities between image elements for scene representation due to gradual variations of illuminations, textures or clutters. To tackle this problem, we utilise a path‐based bottleneck analysis method that captures the semantic information between image elements to measure the dissimilarity. By integrating both the spatial continuity and feature consistency into the understanding of the semantic information, we detect the bottlenecks on the proposed double‐S path to define the bottleneck distance, which demonstrates a favourable capability of grouping image elements that follow a similar pattern and separating different ones. In the experiments, the method is proved to be robust to noises and invariant to changing illumination and arbitrary scales in natural images. Tests on some challenging datasets validate the advantage of applying the path‐based bottleneck distance in image ranking and salient object detection.

The authors propose a tracking algorithm based on the reliability analysis of the convolutional neural network to avoid drift. In general, most tracking algorithms implemented with the deep network consist of a single network; they obtain the tracking results according to the confidence and perform updates with the samples, which are collected based on the previous target state. However, this kind of algorithm relies heavily on the accuracy of tracking results, and slight deviations can lead to improperly labelled training samples and degrade the network. Therefore, they design a verification network to guarantee the reliability of the tracking network by correcting the results and it can be connected to a tracking network by sharing convolutional layers. The reliability verification network estimates the accuracy of the results of the tracking network and discards ambiguous results to avoid accumulating errors. Specifically, the verification network can distinguish the target from the confused candidates more precisely because of the optimised training data. The training samples of the verification network consist of characteristics and labels, and they are optimised by feature selection and label enhancement, respectively. The experimental results illustrate the outstanding performance compared with several state‐of‐the‐art methods on the challenging video sequences.

Local saliency of a target describes the difference between the target and its surroundings, an effective way to improve the efficiency of searching for a target is to enhance its local saliency. To resolve the visual search and localization problem in vision-based robotics, this paper proposes a stepwise method that approaches the target in a successive way. The method has two main features: (1) a search path is established in searching for objects with high saliency based on saliency analysis; (2) In the process of target search and localization, the relationship between object locations is utilized to reduce the size of search regions step by step and, simultaneously, enhance the saliency of objects inside regions. Experimental results indicate the efficiency of the method for target search and localization under complex scenes.