Research On High-efficiency Target Detection Algorithm Based On Embedded Platform

Deep learning relies on its advantages of plasticity and universality to demonstrate its outstanding performance in computer vision and other fields.In computer vision,object detection algorithms have been widely used in several fields such as unmanned driving and pedestrian monitoring.As the accuracy of the object detection algorithm increases,the amount of parameters and calculations of the algorithm increase exponentially.The deep learning-based object detection algorithms need to run on a powerful computing platform,and cannot be directly applied to embedded devices.While mobile devices and smart terminal devices are the mainstream of human life,deep learning models are difficult to run in mainstream devices.To address the above limitations,this paper proposes a lightweight and efficient object detection algorithm based on deep learning to achieve real-time detection on multiple embedded devices.For mobile embedded devices,this paper is based on a single-stage YOLO network.In the residual module,our method uses deep separable convolution to replace the traditional convolutional layer,integrates feature scale regularization strategy,and incorporates the random deletion of level dependencies strategy.While reducing the overall calculation amount of the model,the accuracy and feature dimensions of the original model are maintained,which significantly improves the speed of the model.Then,the loss function is improved based on the coincidence degree and relative distance of the detection frame,and the maximum value suppression method is constrained by the frame loss,so that the model learns the spatial relationship between the prediction frame and the real frame faster,softens the prediction frame elimination algorithm,and speeds up the speed of model convergence.The extensive experimental results demonstrates that our method has achieved 73.26% and 84.58% accuracy on the Pascal VOC and flame detection data sets,respectively,and runs in real-time on the mobile embedded system at a speed of 47 FPS.For the Jetson TX2 embedded devices,based on the Center Net high-efficiency object detection network,this paper designs a multi-level and multi-dimensional information attention feature weighting mechanism to generate pixel-level attention focus on the model's receptive field,so that the model can distinguish the object from the background.In the data processing stage,this paper introduces a high-efficiency data enhancement strategy to enhance the image source,image background,and image mixed noise,so that the diversity and complexity of the data and the robustness of the model can be well improved.The extensive experimental results demonstrates that our method has achieved 80.71% accuracy in the Pascal VOC object detection data set,and it runs in real time on the common embedded system at a speed of 27.4FPS.