Research Of The High Efficiency Infrared Images Pedestrian Detection Algorithm Based On Deeply Supervised Learning

Pedestrian detection has always been an important research topic in the field of computer vision.Real-time and accurate pedestrian detection systems are required in areas such as driverless,robot navigation,security,and intelligent transportation.Due to the imaging defects of visible light images under poor illumination conditions,far infrared thermal images have become an important source for all-weather pedestrian detection based on visual methods.In recent years,with the rise of deep learning technology in the field of computer vision,the use of convolutional neural networks for pedestrian detection has attracted the attention of many researchers,and has achieved outstanding performance in some areas.Large-scale data is crucial for the effect of deep learning,but the available labeled infrared data is much less than the common visible light data.Most of the current infrared pedestrian detection methods are pre-trained in the ImageNet dataset,and then finetuned on infrared data sets.However,the far-infrared image is a single-channel grayscale image,and the imaging details are also different from the visible light image.In addition,the classification model turns to the migration learning bottleneck of the detection task.These factors greatly affect the accuracy of neural network training results and limit the performance of infrared pedestrian detection system.In practical applications,larger and more complex neural network models have been proven to be more effective and widely used in products,which also has greater requirements and consumption for computing resource and storage bandwidth.The main research work of this thesis is as follows:1.We proposed a pedestrian detection algorithm based on deep supervised learning.Add dense residual connections in the backbone sub-network and back-end prediction generation sub-network of the network to improve information flow and feature reuse between different layers of the network.At the same time,the feature extraction process is optimized for the target detection and positioning tasks,Finally,the pedestrian detection network can realize the target that the network is not based on the pre-training model on visible image dataset and directly training from scratch on the infrared image data set,and the detection precision is higher.2.Optimize the feature extraction process for target detection and positioning tasks,replace the traditional convolution with depth separable convolution in the predictive generation sub-network,and select the large-resolution feature map to improve the detection accuracy of large targets and reduce the leakage of small targets.Check rate,and network model size and computational overhead are not increased3.The pedestrian detection algorithm is optimized for a single instruction multiple data architecture.Based on the micro-architecture and related parameters of the hardware platform,the optimization strategy is formulated,and the core operator depth separable convolution is rewritten and optimized.Take measures such as data layout transformation and loop reordering,register partitioning,cache blocking,etc.,and write vectorization-friendly and cache-friendly code.The optimized core operator's actual running time is significantly reduced,so that the target detection network of this design can be run in real time on the hardware platform.