A multi-target detection method for vehicles and pedestrians based on improved SSD network

Abstract

The invention discloses a vehicle and pedestrian multi-target detection method based on an improved SSD network. Step S1: a magnanimous driving video collected by a driving recorder is made into a picture input set with an appropriate size; Step S2: According to k-Means clustering method modifies the size and aspect ratio of the anchor frame in SSD network to fit the data set; Step S3, replacing the original loss function with the Focal Loss function, replacing the original Online Hard Example Mining (OHEM) mechanism, and solving the problem of positive and negative sample imbalance; Step S4,inputting a data set and training a new SSD network; Step S5: Using the trained SSD network to detect the object of the real-time input picture. According to the technical proposal of the invention, by means of clustering, Set the size and aspect ratio of the proposed anchor frame, By using Focal Loss function, the problem of the imbalance between positive and negative samples is solved, the importance of difficult-to-separate samples is increased, the occupation of memory is reduced, the training speed is improved, and the accuracy of the whole detection is also improved.

Description

technical field [0001] The invention belongs to the field of target detection of computer vision, can be applied to the fields of unmanned driving, safety monitoring, road supervision, etc., and in particular relates to an improved single shot multibox detector (Single Shot multiboxDetector, SSD) network-based detection of vehicles, pedestrians, etc. Multi-object detection method. Background technique [0002] With the development of science and technology, target detection has become a hot research direction of computer vision, which can be applied to unmanned driving, video surveillance, pedestrian detection, sea ship detection and other fields. In the past, traditional machine learning methods were basically used for target detection, that is, operators (such as: HOG, SHIFT, Haar) were used to extract features, and then classifiers (such as: SVM, Fisher, Adaboosting) were selected to perform these features. Classify to get the result of target detection. But the traditi...

AI Technical Summary

Problems solved by technology

However, there are two fatal shortcomings in the traditional method: First, the selected operator will not be suitable for the extraction of all features, that is, some of the extracted features are not very representative

[0005] 1. The detection effect on small targets is poor, and only 300*300 and 512*512 models are given, that is, the original model is only suitable for images with resolutions of 300*300 and 512*512, and cannot be input with high-resolution images to increase resolution of small objects

[0006] 2. When dealing with the imbalance of positive and negative samples, additional computing time and storage space are required, and the influence of easily divided samples on network convergence is completely ignored.

Images