Traffic person and vehicle non-target detection method for small target

Abstract

The invention provides a traffic person and vehicle non-target detection method for a small target. The method comprises the following steps: firstly, obtaining a to-be-trained traffic image sample set; performing data enhancement on the data of the small target; then constructing a small target detection network based on the yolov5, inputting the image sample set with the label into a neural network, updating a parameter weight of the neural network according to a partial derivative of a loss value to a network parameter, and when the loss value is minimum, obtaining a trained human-vehicle non-target detection model for a small target; and finally, inputting a to-be-detected traffic picture into the detection model, and outputting the category, confidence and four position coordinates of each target in the image. According to the invention, by using a data enhancement method for small targets and constructing a feature fusion module based on an attention mechanism, real-time detection of non-targets of traffic people and vehicles is realized based on a YOLOV5 detection network, and the detection rate and the detection accuracy of the small targets are greatly improved.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a non-target detection method for traffic people and vehicles aimed at small targets. Background technique [0002] With the advancement of urban intelligence and digitization, intelligent transportation has become an indispensable link in smart cities. The development of intelligent transportation not only saves the manpower required for traffic supervision and inspection, but also solves the problem of supervision in harsh traffic environments. It is difficult to detect the occurrence of traffic incidents more comprehensively and timely, which brings great convenience for people to travel. The detection of pedestrians, vehicles and non-motor vehicles in road traffic can help improve the efficiency of traffic supervision, reduce traffic congestion, reduce the probability of traffic accidents, and reduce the manpower required for criminal investigation. However, due to ...

AI Technical Summary

Problems solved by technology

[0006] 1) There are relatively few pictures containing small targets in the data set, which causes the model to be biased towards medium or large targets during training; the area of ​​small targets is too small, resulting in fewer anchors containing targets, which also means that small targets are The probability of detection becomes smaller

[0007] 2) Excessive downsampling rate and receptive field: Assume that the current small object size is 15×15, and the convolution downsampling rate in general object detection is 16, so that on the feature map, the small object does not even occupy a single point ; and the receptive field of the feature points on the feature map is much larger than the downsampling rate, resulting in a point on the feature map, small objects will occupy fewer features, and will contain a large number of features in the surrounding area, thus affecting its detection effect

[0008] 3) The existing general-purpose deep convolutional neural network algorithm is easy to lose the characteristics of small targets

Images