To address the issues of low accuracy in existing small object detection algorithms, an improved network model algorithm called YOLOv5s_2E is proposed. This method first uses the k-means++ clustering algorithm to calculate the prior boxes of the Visdrone dataset. Then, it introduces Soft_NMS and combines it with EIoU to propose the EIoU_Soft_NMS algorithm to replace the non-maximum suppression (NMS) of the original network, improving the detection of objects that are occluded. The bounding box regression loss function uses Focal-EIoU, which speeds up model convergence and reduces loss. Additionally, a detection layer is added to the original detection head to unify the channel numbers, and with the dynamic head framework DyHead, the attention mechanism is integrated with the detector’s head to further improve small object detection accuracy. Finally, the system robustness is improved by adjusting the ratio of data augmentation methods Mixup and Mosaic.Experimental results show that the proposed algorithm improves the mAP@0.5, mAP@0.5:0.95 and detection accuracy by 12.6%, 12.2%, and 20.5%, respectively, compared to the previous method on the VisDrone dataset. The parameter size only increases by 4%, and the weight file size increases by only 0.57MB, meeting the accuracy requirements for small object detection.

Aiming at the problems of the YOLOv5, such as large model size, enormous amount of computation, and low accuracy of target box regression, a new model with fewer parameters, less computation, faster convergence speed, and higher accuracy was proposed. Firstly, the improved SKConv was used to greatly reduce the number of model parameters and increase the receptive field range of the network. Secondly, C-ECA, an optimized version of the channel attention module ECA, was added to the model to obtain attention information in a cross-channel way, so that the model could more accurately focus on important features among complex features. Then, the MSM structure is designed to effectively improve the feature extraction ability of the network. Finally, the backbone network is deepened to extract more intermediate features, and the depth of the feature pyramid and detection layer are increased accordingly so that the network can make full use of intermediate features and accurately detect more targets. The experimental results show that compared with YOLOv5s, the number of parameters of the final models is reduced by 18.6%, and the calculation amount is reduced by 8.1%, and when tested on the VOC2007 test dataset, mAP@0.5 and mAP@0.5:0.95 improved by 3.3 percentage points and 5.6 percentage points, respectively. The improved model has significantly improved the ability of target detection in various complex environments and realizes the target detection model with low parameters and high accuracy, which provides important ideas for the design of lightweight target detection networks.

Siamese network based trackers describe object tracking as a similarity matching problem and these trackers achieve state-of-the-art performance on multiple benchmarks. However, due to the non-update of the appearance template and the change of the object appearance, the tracking drift problem often occurs, especially in the background clutter scene. Effective appearance template update methods can improve tracker performance, but most trackers use simple linear interpolation to update the template or do not update the initial template at all. Through experiments, we find that the channel response of the search region with the adjacent frame appearance template is often better than that with the initial frame appearance template. So we add the results of the previous frame prediction as a new template branch to the Siamese network to form a Triplet network. We applied the Triplet network to the SiamFC and SiamCAR, called TripFC and TripCAR. We tested on four challenging benchmarks (GOT-10K, OTB2013, OTB2015, UAV123). The experiments show that our method is powerful and effective, it can be easily embedded into the Siamese trackers. TripFC has a good effect on solving the problem of tracking drift. If necessary we can publish the code to facilitate research in this area.

Aiming at the problems of occlusion, drift, and background change in target tracking, a background learning correlation filtering algorithm based on multi-feature fusion is proposed. In the framework of correlation filtering, multi-feature fusion, multi-template update, and background learning regularization are used to improve the performance of the filter in the problem of template contamination and object occlusion. The fast directional gradient histogram (FHOG), color feature (CN), and texture feature (ULBP) were extracted, and the feature channels were connected in series. Then the depth features of Conv4-4 and Conv5-4 layers were extracted through the VGG-19 network, and the appearance model of the target was constructed. To reduce the sensitivity of the filter to the sudden change of background, a background learning filter is constructed, and the alternate direction multiplier method (ADMM) is used to speed up the calculation of the filter. In the model update stage, aiming at the problem of pollution of the original template caused by target occlusion, a high-confidence multi-template fusion update strategy is proposed by fusing the template with the highest confidence in the current frame, the previous frame, and the history frame. Finally, the proposed algorithm is tested on OTB50, OTB100, UAV123, and TC128 experimental data sets, and some classical and latest algorithms. The experimental results show that the tracking accuracy and robustness of the correlation filtering algorithm are improved.