Research On Real-time One-class Object Detection

The object detection as a relatively mature computer vision technology,is play-ing an increasingly important role in practical applications such as surveillance and localization,traffic and defense.The neural network models for multi-category object detection problem,which develop from the initial RCNN series to the later SSD and YOLO,have received a lot of academic attention and their detection accuracy is grad-ually increasing.However,the neural networks are not required to detect too many categories in many practical scenarios,and scenarios where only one category needs to be identified are common.Thus,the need for real-time detection is more important.If the above-mentioned multi-category object detection models are adopted directly,the detection cannot be completed in real time on ordinary arithmetic computers for the networks have too many operational parameters,which brings difficulties for practical applications.For the problem of real-time one-class object detection,we propose two methods to reduce current model parameters to speed up their computation,and build an indoor localization system based on the above optimization algorithms.The main works of this paper are as follows:1.For achieving real-time one-class object detection,we propose an optimization strategy for the feature extraction base model mobilenet v3,modifying the imple-mentation details of the SE(Squeezeand Excitation)layers and the connection se-quence of depthwise,SE,and pointwise layers.We finally obtain a new network model mobilenet v3 small₂with less number of parameters and higher accuracy than the original mobilenet v3 combining with the warmup training method.It is used as the backbone model of the SSD network in following experiments for the following real-time one-class object detection,which has obvious speed advantage compared with other network models.2.To further improve the precision and speed of the above model,we propose a network pruning algorithm for one-class object detection,which adaptively se-lects different convolutional kernels for culling according to the object categories.Also,we analyze the pruning strategy from the viewpoint of interpretability.It is demonstrated in experiments that the pruned network of this algorithm has certain advantages in terms of speed and accuracy of object category detection compared with the pre-pruned network.3.Based on the above two methods,an one-class(person)object detection model is trained and an indoor localization system is built on it.The system can achieve real-time detection and high detection accuracy on a computer with ordinary computing power.Relevant experiments show that the optimization algorithm and pruning strategy proposed in this paper can effectively improve the computation speed and accuracy of object detection network for one-class detection.Meanwhile,the indoor localization system proves the practicality and feasibility of the above algorithms.