After obtaining the first place in the open dataset ranking in 2018 and the large-scale person search competition of PRCV2018, Dahua AI Person Re-identification technology recently made another breakthrough – in the three international authoritative open data sets Market1501, DukeMTMC-reid and CUHK03, the key indicator mAP reached 91.98%, 83.96%, and 85.72% respectively, refreshing the best performance record and continuing the leading position in algorithm. This year, Dahua AI technology has acquired the top rankings in many international competitions fields such as semantic segmentation and instance segmentation etc.
Person re-identification technology, referred to as Person-ReID, uses computer vision technology to retrieve the same target person under different cameras. This technology has great practical value for carrying out artificial intelligence businesses because of the difficulty to recognize the same target due to the different image shooting angles, resolution, target postures, obstacle occlusion, and uneven illumination under different cameras. Based on years of technical accumulation in vehicle and person recognition, Dahua Technology has achieved breakthroughs in the following three areas to facilitate the AI application.
First of all, Dahua Technology’s innovative application of image data enhancement methods, including random blur and random interception strategies, enabled the effective simulation of complex situations such as body occlusion, blur and incompleteness in various environments. The random interception strategy is conductive to mine the potential feature extraction of the block components network and to improve the network feature matching performance.
Secondly, for the problem of feature granularity difference in multi-branch component network, a progressive part model (PPM) is adopted. There is cascaded semantic relation among the branches besides the shared basic convolution network.
Finally, while designing PPM network, the overlapping sampling operation is used to facilitate the extraction of more striking feature information from each component branch, and the improved loss function is used to learn the feature embedding space based on spherical constraints. In addition, a branch based on the attention scoring mechanism is added to the PPM feature layer to enable the network to adaptively combine the multi-granularity features of the human body from each branch.
The person re-identification technology has been successfully applied in Dahua video structured cameras and video structured analysis servers, and has maintained a leading place in the search by image performance of ultra-largescale pedestrians, non-motor vehicles and motor vehicles, targeting smart cities, buildings, supermarkets and other places to reduce user costs, improve work efficiency and video development application value.
