Research On Object Tracking Algorithms Based On Model Lightweighting

The continuous enhancement of CNN in last several years.The computational complexity of tracking algorithms for inference has become immense,which cannot be adapted to real life.And the object tracking is affected by background clutter and drastic changes in illumination during the tracking process.In view of this phenomenon,this paper investigates tracking technology.It need realize the lightweight of object detection algorithm.At the same time,it needs to improve object tracking accuracy.The main work completed includes the following:(1)To address the problems of computationally intensive and complex models in object detection algorithms,which cause difficulties in deployment in application scenarios with limited computational resources on embedded platforms.In this paper,a lightweight object detection method based on feature fusion with an optimised Retina Net model is proposed.Firstly,the feature of low number of depthwise separable convolutional parameters is exploited to build a stacked new compact network for feature extraction in order to reduce the number of model parameters.The scale invariance of the features is improved by a spatial feature fusion mechanism.Secondly,the idea of structural reparameterisation is incorporated to increase the training depth to achieve multi-branch training and single-branch inference,which better improves the inference performance of the model.The experimental results show that the average accuracy of the optimised object detection algorithm is 54.1%,which is higher than the average accuracy of Retina Net;the memory occupied during inference is 44.27% of the memory occupied by Retina Net,which greatly improves the inference speed of the network while ensuring accuracy.(2)To address the more difficult problems of image deformation and low resolution during fast motion object tracking.Based on current Siamese network,a featureadaptive fusion object tracking algorithm based on a normalized attention mechanism is proposed.Less obvious weights are suppressed by a lightweight attention mechanism.A path enhancement method is also proposed,whereby path enhancement is applied to the last four feature layers of the backbone network.The experimental results show that the proposed algorithm achieves a high success rate of 59.5% in the constructed dataset,outperforming the mainstream trackers.At the same time,the algorithm has strong robustness when the external lighting changes,the image background is complex and the object is rotated in-plane.(3)A target detection and tracking system is designed to test the target tracking effect in the actual scene.The experiment in the actual application scenario verifies that the tracking effect is good under the circumstance of complex background environment and intense illumination change,which proves the applicability of the optimized algorithm and design system.