Research On Object Detection Algorithms For Complex Environments

Object detection is the pillar technology of computer vision and image processing.Object detection algorithms identify the specific areas from the input image in the form of a bounding box and category.Object detection is not only a hot research topic,but also the foundation of higher-level computer vision tasks such as scene understanding,object tracking,image description,and event detection.With the development of deep learning algorithms based on convolutional neural networks,object detection algorithms developed rapidly.Object detection algorithms have been successfully applied in many fields such as autonomous driving,intelligent monitoring,human-computer interaction,intelligent medical care,intelligent robots,and augmented reality.Although object detection algorithms developed fast and have excellent achievements.There are still many limitations and challenges in various complex environments.For instance,there are:1.Edge environments with limited computing resources.Large backbone network becomes more and more complicated in order to improve measurements of object detection benchmark datasets.The edge environment cannot work under the heavy load from these complicated algorithms.Replacement with common lightweight backbone networks often results in sharp performance degradation.2.Complex environments with special objects.Input images from complex environments tend to be quite different from those in benchmark datasets.The objects to be detected from actual sites often have some special properties.Applying the general algorithm with the default structure and configuration mechanically will cause problems of incompatibility.In the face of the difficulties and problems mentioned,this paper selects typical complex environments including the edge environment with limited computing resources and the aerial perspective environment containing dense small objects.The main contributions made by this paper are as follows:1.For the edge environment,this paper proposed a new lightweight anchor-free detector by the name of Cross Net.Firstly,this paper analyzed the defects in the baseline model Center Net,and then replaced the heavy backbone network and removed the unconventional operator.Large receptive field convolution layers and information fusion are introduced into the network structure for better feature reconstruction in the next step.Secondly,the enlarged softening and the novel sample matching strategy based on neighbor points are introduced in the training phase.And the dynamic weighted regression loss is also used.Cross Net maintains the same accuracy compared with the baseline model,while it achieved a ten-fold improvement in computation and parameters.2.For the complex environment with special objects,this paper proposed a high-performance detector Air-YOLO for dense small objects.First of all,this paper analyzes and visualizes special objects,and points out the conflict between dense small objects and network design with quantitative indicators.Then,a new cross-scale split feature neck is proposed to match the best scale choice with less computation and parameters and bring knowledge distillation to improve the model performance.Air-YOLO maintains the same amount of computation and parameters compared with the baseline model,and the detection performance is greatly improved.The research findings of this paper solve the problems caused by the two complex environments partly,which have great application value.At the same time,the analysis and improvement process in this paper has certain enlightening significance for similar scenarios and other algorithms.