Dense pedestrian detection method

Abstract

The invention discloses a dense pedestrian detection method. In the network training stage, a convolutional neural network is built, an encoder and a decoder are trained, and a pedestrian bounding box with correct head prediction is detected; according to a prediction result, an optimal prediction candidate frame is distributed to each pedestrian; other network parameters except the density estimation head are fixed, the unique prediction candidate box of each pedestrian is used to generate a density target, and the density estimation head is trained; and finally releasing all parameters of the network, and jointly training the whole network. And in a test application stage, when post-processing is carried out, subtracting the Gaussian activation map at the corresponding position from the predicted density map every time a pedestrian frame which is determined to be reserved is selected, and carrying out secondary judgment on the pedestrian frames of which the overlapping rate with the reserved pedestrian is greater than a threshold value by utilizing the updated density map. Under the dense scene, the problem that a general non-maximum suppression method mistakenly deletes a correctly predicted bounding box is solved, and meanwhile, the performance of the non-dense scene is not influenced.

Description

technical field [0001] The invention belongs to the technical field of image pedestrian detection, and is used to solve the problem of a low overall recall rate caused by wrongly deleting correctly predicted candidate frames in the case of dense pedestrians in a non-maximum value suppression method commonly used in the post-processing stage of a detector. Background technique [0002] Pedestrian detection is a very challenging computer vision task, and it is widely used as a core module in various computer vision systems. Although pedestrian detection technology has made significant progress in recent years, pedestrian detection in dense conditions is still a very challenging task due to frequent occlusions in practical applications. [0003] In recent years, methods based on convolutional neural networks have achieved absolute dominance in the field of pedestrian detection, and the accuracy is much higher than that obtained by using manually designed features. Pedestrian d...

AI Technical Summary

Problems solved by technology

This method has made a good improvement, but this method is still different from the human discrimination mechanism. People can distinguish whether two pedestrians are the same pedestrian without comparing the similarity. In the scene where the pedestrians look similar ( such as in a campus or a factory where people wear the same clothes) this approach may fail

Images