Robust Visual Tracking via Constrained Multi-Kernel Correlation Filters

Abstract

Discriminative Correlation Filter (DCF) based trackers are quite efficient in tracking objects by exploiting the circulant structure. The kernel trick further improves the performance of such trackers. The unwanted boundary effects, however, are difficult to solve in the kernelized correlation models. In this paper, we propose a novel Constrained Multi-Kernel Correlation tracking Filter (CMKCF), which applies spatial constraints to address this drawback. We build the multi-kernel models for multi-channel features with three different attributes, and then employ a spatial cropping operator on the semi-kernel matrix to address the boundary effects. For the constrained optimization solution, we develop an Alternating Direction Method of Multipliers (ADMM) based algorithm to learn our multi-kernel filters efficiently in the frequency domain. In particular, we suggest an adaptive updating mechanism by exploiting the feedback from high-confidence tracking results to avoid corruption in the model. Extensive experimental results demonstrate that the proposed method performs favorably on OTB-2013, OTB-2015, VOT-2016 and VOT-2018 dataset against several state-of-the-art methods.

Improving Object Tracking by Added Noise and Channel Attention

“…We observe that CCOT [ 40 ] secured best EAO 0.33 but our IRCA-Siam algorithm showed better accuracy and robustness for VOT2016 dataset. CMKCF [ 72 ] have shown lower robustness compared to our method but its accuracy is lower than ours. Moreover, our method showed best accuracy 0.56 compared to other state-of-the-art methods for VOT2016.…”

“…We compared our method with 30 state-of-the-art trackers including SiamTri [ 55 ], CSRDCF [ 48 ], CNNSI [ 56 ], SRDCF [ 57 ], Staple [ 58 ], TRACA [ 59 ], SiameseFC [ 15 ], CFNet [ 10 ], ACFN [ 24 ], SiamFc-lu [ 60 ], HASiam [ 61 ], SiamFCRes22 [ 62 ], Kuai et al [ 63 ], MSN [ 64 ], MLT [ 65 ], KCF [ 66 ], SCT [ 67 ], OA-LSTM [ 68 ], ECOhc [ 23 ], DSiam [ 69 ], MEEM [ 70 ], CCOT [ 40 ], SAMF [ 71 ], CMKCF [ 72 ], SATIN [ 73 ], GradNet [ 74 ], SiameseRPN [ 75 ] DSST [ 76 ], MemTrack [ 14 ], MemDTC [ 77 ], and UDT [ 78 ].…”

“…We present the performance comparison over VOT2016 and VOT2017 in Table 4 and Table 5 respectively. We compared our method over VOT2016 with various state-of-the-art trackers such as MemTrack [ 14 ], MemDTC [ 77 ], ECO [ 23 ], HASiam [ 61 ], Staple [ 58 ], SRDCF [ 57 ], DSiam [ 69 ], MLT [ 65 ], CCOT [ 40 ], UDT [ 78 ], SiameseFC [ 15 ], CMKCF [ 72 ], and SiamFCRes22 [ 62 ]. We observe that CCOT [ 40 ] secured best EAO 0.33 but our IRCA-Siam algorithm showed better accuracy and robustness for VOT2016 dataset.…”

Learning Soft Mask Based Feature Fusion with Channel and Spatial Attention for Robust Visual Object Tracking

“…We evaluate the proposed tracking versions on five benchmark datasets including OTB2015 [ 44 ], TempleColor128 [ 45 ], UAV123 [ 46 ], VOT2016 [ 47 ], and VOT2017 [ 48 ]. We compare its performance with 39 state-of-the-art methods including TRACA [ 71 ], SRDCF [ 72 ], UDT [ 73 ], SiamtTri [ 31 ], SiameseFC [ 30 ], Staple [ 74 ], CFNet [ 34 ], CNNSI [ 75 ], RNN [ 76 ], MLT [ 77 ], P2FNet [ 78 ], SiamFc-lu [ 79 ], DSiamM [ 57 ], Li et al [ 80 ], Kuai et al [ 81 ], SINT [ 82 ], HASiam [ 83 ], ECO [ 21 ], MEEM [ 84 ], SAMF [ 85 ], MUSTER [ 86 ], CSK [ 6 ], CMKCF [ 87 ], SATIN [ 88 ], ACT [ 89 ], MemTrack [ 58 ], DSiam [ 90 ], GradNet [ 91 ], SiameseRPN [ 92 ], ACFN [ 93 ], CSRDCF [ 68 ], SCT [ 94 ], KCF [ 7 ], SSKCF [ 47 ], DPT [ 95 ], DSST [ 96 ], CCOT [ 20 ], SiamDCF [ 97 ], and UCT [ 98 ].…”

“…Table 6 presents the Expected Average Overlap (EAO), accuracy (A), and Robustness (R) of the compared trackers over VOT2016 [ 47 ] for the baseline experiments. Our methods are compared with 11 trackers: MemTrack [ 58 ], MemDTC [ 90 ], ECO [ 21 ], SRDCF [ 72 ], DSiam [ 57 ], CMKCF [ 87 ], Staple [ 74 ], CCOT [ 20 ], UDT [ 73 ], and SiameseFC [ 30 ]. CCOT obtained the maximum EAO score, and its robustness value is larger than our proposed algorithm, and overlap scores is less than our algorithm, as presented in Table 6 .…”