Research Of Object Detection Based On Deep Learning And Improved Target Tracking Algorithm Based On KCF

Nowadays,the research of target detection and target tracking have become hotspot in the field of computer vision,meanwhile,widely used in transportation,medical and military fields,intelligent people’s life.In practical engineering application,the detection algorithm is demanded to be high accuracy and real-time,in addition,the target tracking algorithm has the requirement of robustness to occlusion and scale variations,so it has vital engineering application value to study the target detection and the target tracking technology.In this paper,the requirements of high-speed and lightweight in practical engineering applications for the detection algorithm are fully considered,and the detection performance of existing algorithms in the detection field is compared and analyzed.Then,we select,study and improve YOLOv4-tiny target detection algorithm,and a light LD＿CFSM detection network is proposed based on the selected algorithm.In order to reduce the number of calculation parameters,the improved network selection depthwise separable convolution,replace the traditional convolution in Backbone network.Meanwhile,CBAM attention mechanism is introduced after the effective feature layer of Backbone network to pay more attention to the channel and space features,so as to maintain algorithm accuracy while lightweight model.Then,a Focus construction is selected to replace the first convolutional module of the main feature extraction network to integrate the information of the original image into the channel space through slice operation to prevent features loss and make the network pay more attention to the learning of target contour features.Then,a SPP-Net structure is introduced before the features enter the Neck part of the network for high and low dimensional fusion,and the results of maximum pooling with different kernel sizes are arranged and combined to expand the perception area of the network.Therefore,the extracted feature reuse has been strengthened and the ability of detection has been enhanced.Finally,we use data enhancement strategies,Rotation and Mosaic,to enrich the training data,improve the ability of generalization and the performance of detection.In this paper,the proposed LD＿CFSM algorithm increases m AP by 9.67% and reduces model storage by 2/3compared with the original algorithm.What’s more,the implemented experiments and analyses prove that the improved LD＿CFSM algorithm is superior to the YOLOv4-tiny algorithm in detection performance and model storage.Following the joint research of intelligent technology,tracking algorithm of correlation filtering is notability for its superiorities of both precision and speed,so the classical KCF algorithm is chosen for further improvement to satisfy the engineering requirements.In this paper,we study the principle of KCF tracking algorithm above all,then optimize the weaknesses of the KCF algorithm,finally gain the modified KCF＿LD tracking algorithm.Specifically,the proposed algorithm adopts multi-scale detection to flexibly adapt to the scale variations of the target by the using of two-dimensional position filter and onedimensional scale filter together.Meanwhile,33 different scales are used for multi-scale detection to increase the sensitivity of the algorithm and improve the performance of tracking.Since KCF will update the model after tracking of each frame,resulting in poor tracking performance in the scenario of occlusion.Therefore,we introduce a reliability mechanism of model update into the detection algorithm for improving the anti-occlusion performance,where the APCE value of the response graph of tracking results is regarded as the criterion of model update.Formally,when the APCE value is lower than the mean of previous values,the tracking will stop and the proposed LD＿CFSM algorithm is then used to start global detection to estimate the location of the target.Meanwhile,the detection accuracy is guaranteed by the detection confidence threshold and resume the tracking tasks.By implementing simulations,we can prove that the proposed KCF＿LD tracking algorithm in this paper has 3.2% improvement in the tracking accuracy and 4.1% improvement in the success rate compared with KCF,showing significant robustness.