Classifying Global Scene Context for On-line Multiple Tracker Selection

Abstract

In this paper, we present a novel framework for combining independent on-line trackers using visual scene context. The aim of our method is to decide automatically at each point in time which specific tracking algorithm works best under the given scene or acquisition conditions.In the literature, many ways of combining, fusing or selecting visual features have been presented. For example, low-level fusion of features (like motion or shape) is applied to improve the foreground-background discrimination (e.g. [2, 8]). Fusion is also possible at a higher level, where several trackers are run in parallel in order to select or combine their respective results (e.g. [1,5]). In terms of model or feature fusion, our previous work Moujtahid et al.[6] concentrated on using confidence values of several individual trackers with different visual features coupled with a spatial-temporal coherence criteria to select the most suitable tracker at a given instant and enforce the continuity of tracking.The main idea behind our framework is to use the strengths of different tracking algorithms as well as scene context information in order to improve the tracking performance. To this end, we introduce a framework that combines several independent and complementary trackers, each specialised on different scene conditions. The decision on which tracker to select is proposed by an off-line trained classifier which, in turn, is based on general scene context features that are independent from the trackers.The general procedure of the proposed tracking framework is illustrated in Fig. 1. On a given video, N independent trackers T n , (n ∈ 1..N) run in parallel and, at every frame t, produce each an estimate of the object's state. This is usually a bounding box B n t with an associated confidence value c t,n . The objective is to select at each frame the best tracker, i.e. the one that outputs the bounding box that fits best the object to track.At the same time, the scene context features f t are extracted. These features correspond to first and second order statistics of a given imagerelated variable such as intensity, colour and motion. They are computed on different image regions giving local, global and differential values.The scene features f t are concatenated with additional measures like the trackers confidences values c t and the identifier of the last selected tracker s t−1 to form a large feature vector i t . An N-class classifier, that has been trained off-line on annotated data, is then applied on these features to estimate the best tracker for the given scene context. The classifier responds with y t , a probability for each class which is subsequently filtered by a Hidden Markov Model to ensure the temporal continuity of the tracker selection and reject outliers. The HMM estimates the posterior probability distribution x t , which is used to select the best tracker.Finally, a Kalman Filter is applied as a post-processing step to temporally smooth the resulting object bounding box B s t from the selected trackers T s . The r...

Visual Tracking Based on Complementary Learners with Distractor Handling

“…Using the VOT2015 benchmark dataset, the proposed LPMT method was compared with the state-of-the-art visual tracking algorithms: ACT [31], CT [35], ggt [15], L1APG [39], mkcf_plus [15], RobStruck [15], STC [40], amt [15], DAT [17], HMMTxD [15], LGT [41], muster [42], s3Tracker [15], sumshift [43], AOGTracker [15], DFT [15], HT [15], loft_lite [15], mvcft [15], samf [21], TGPR [44], ASMS [45], DSST [16], IVT [25], LT_FLO [26], ncc [27], SCBT [46], tric [15], dtracker [15], kcf_mtsa [15], matflow [15], OAB [15], sKCF [15], zhang [15], bdf [15], fct [15], KCF2 [15], MCT [15], OACF [11], sme [15], cmil [15], fot [33], kcfdp [15], MEEM [47], PKLTF [15], SODLT [48], CMT …”

Convolutional Shallow Features for Performance Improvement of Histogram of Oriented Gradients in Visual Object Tracking

“…IVT [24], kcf mtsa [19], KCF2 [19], kcfdp [19], kcfv2 [19], L1APG [9], LGT [25], loft lite [19], LT FLO [26], LPMT [15], matflow [19], MCT [19], MEEM [27], MIL [4], mkcf plus [19], muster [28], mvcft [19], ncc [29], OAB [19], OACF [19], PKLTF [19], rajssc [19], RobStruck [19], s3Tracker [19], samf [30], SCBT [31], sKCF [19], sme [19], SODLT [32], srat [19], STC [33], struck [34], sumshift [35], TGPR [36], tric [19], and zhang [19]. Further, to prove the advantages of the proposed method, a comparison among the proposed method, the proposed method using only CS features, and the proposed method using only HOG features was also conducted.…”

Beyond Background-Aware Correlation Filters: Adaptive Context Modeling by Hand-Crafted and Deep RGB Features for Visual Tracking

“…The proposed method is extensively compared with state-of-the-art visual trackers on the OTB-50, OTB-100, TC-128, UAV-123, and VOT-2015 dataset, which their benchmark results are publicly available. These visual trackers include KCF [22], SRDCF [10], HCFT [29], HCFTs [30], LCT [31], LCTdeep [31], SiamFC-3s [3], DCFNet [44], DCFNet-2.0 [44], Staple-CA [35], SAMF-CA [35], BACF [17], DSST [13], Struck [19], MUSTer [23], Staple [2], MDNet [37], DeepSRDCF [11], SODLT [24], C-COT [12], TADT [25], SA-Siam [20], Siam-MCF [33], SAMF [26], MEEM [51], ACT [7], TGPR [18], ASMS [42], OAB [24], SCBT [34], EBT [53], SumShift [24], MIL [1], FCT [24], DAT [39], and sKCF [24]. The OTB-50, OTB-100, TC-128, UAV-123, and VOT-2015 datasets have 51, 100, 129, 123, and 60 challenging video sequences, respectively.…”