Object Ttacking Algorithm Based On Embedded GPU Research And Implementation

Vision-based object tracking is an important basic research direction in the field of computer vision.Many advanced applications based on computer vision require the support of object tracking algorithms such as autopilot,behavior analysis,intelligent monitoring and interpersonal interaction.Vision-based object tracking tasks encounter various problems such as scale changes,shape changes,illumination changes,fast motion,and occlusion.At the same time,it is necessary to ensure that the algorithm has good robustness and real-time performance.So this will be very Challenging puzzles.Based on the algorithm of correlation filtering,this paper designs a tracking algorithm with high accuracy,good robustness,long-term tracking performance and good deployment on embedded platform.The main work and research contents of this paper are as follows:1.Consider the advantages and disadvantages of the DCFNet tracking algorithm.The DCFNet tracking algorithm will be modified.Firstly,the tracking confidence evaluation method is added for the DCFNet tracking algorithm.Then the template update strategy of the relevant filter layer is formulated.The DCFNet tracking algorithm is used as a local tracking module.And a global search module is added.The global search module is started when the confidence of the tracking result of the local tracking module is relatively low.In order to ensure the speed of the global search module,the idea of a cascaded classifier is adopted.First,the Selective Search method is used to filter out the rectangular frame that may contain the target object.Then the three-step screening is performed.The first step,use the size of the object of the previous frame to filter the candidate frame;The second step,use the color features of the target frames of the previous frame and the correlation filtering method to learn the detection template parameters.And use the template parameter to filter the candidate frame;The third step,utilize the DCFNet's convolutional network layer to extract the features of the candidate target.Then use the correlation filter layer of DCFNet to calculate the tracking confidence scores for candiate target objects.Compare these with the result of the tracking module to select the optimal target object.The experimental results show that the improved algorithm has a certain performance improvement compared to the DCFNet tracking algorithm,especially for video frames where moving objects are occasionally occluded.2.Analyze the performance bottleneck of the improved tracking algorithm.On the multi-core embedded development platform Jetson TK1 built by ARM CPU + NVIDIA GPU,realize shared memory-based direct convolution operation and matrix based multi-matrix multiplication for specific convolution operations.The two GPU-based convolution acceleration methods are compared with the convolution acceleration method in cuDNN to select a specific convolution acceleration method that is more suitable for the Jetson TK1 platform.With the ncnn framework,the object detection algorithm based on YOLOv3 is deployed and accelerated on the ARM platform.3.A concrete graphical user program is constructed,which combines object detection,object tracking algorithm and hard acceleration method based on embedded ARM and GPU hardware platform.Realize a complete moving object detection and tracking system.The entire system is divided into a main window module,a target selection module,an image acquisition module,and an image processing module.Two threads are simultaneously opened in the main process for video frame acquisition and processing.