Improving Person Re-Identification Based on Human Height Information

Over the last few decades, threats to urban security have increased worldwide. Surveillance solutions aim to detect, track, and identify people of interest using a large number of surveillance camera videos. In particular, in Japan, the police force expects surveillance solutions to help criminal tracking. As the number of surveillance cameras has increased in recent years, manual analysis has reached its limit, resulting in a high expectation for automatic analysis using artificial intelligence (AI). The AI study on person re-identification has a high potential for human verification in academia; however, it is still challenging to apply. This is because practical tracking accuracy cannot be obtained when a person changes clothes on the assumption that the criminal is on the run. Therefore, we have developed and evaluated an AI technology that identifies people in a scene by focusing on human height, a body characteristic that does not change. To estimate the height from the video, it is necessary to obtain the camera specific parameters. It is difficult to estimate the parameters of many existing cameras because well-known and other conventional camera calibration methods require tedious manual procedures and preliminary information such as objects of known length to estimate the camera parameters. Therefore, we developed a high-precision height estimation method that does not require prior knowledge or the complicated manual procedure by estimating camera parameters using the skeleton of a person captured by a surveillance camera. Moreover, we developed a method that can determine whether a person is the same person even if the person's clothes change using the human height, representing one of the characteristics unique to the person, together with the image features. When evaluating images to identify similarity in people in different clothing captured at size of >200 pixels, the mean average precision improved by 64.8% compared with the conventional method, thus suggesting practicability. In future, we plan to confirm the accomplishment of this study using videos closer to the actual environment and use the results to support crime prevention efforts.