Street theft crime remains a significant public safety concern in China and a central focus of urban crime prevention initiatives. Recent studies have combined street view images and deep learning networks to detect on-street population and streetscape physical environment features, highlighting their significance in comprehending the occurrence of street crime. However, the question of whether on-street population better represents the ‘true risk population of street crime’ compared to previous static (residential) and dynamic (ambient) measures, and whether this advantage persists across different times of the day, remains unexplored. Additionally, the association between streetscape physical environment features and street theft crime, as established in the existing literature, has not been thoroughly examined with respect to temporal variations throughout the day. To address these gaps, this study employed a machine learning model and a simultaneous negative binomial regression model to explore the influence of on-street population and streetscape physical environment on street theft counts. Specifically, we assessed whether their effects exhibit noticeable variations between day and night. Furthermore, we controlled for potential effects of crime attractors, generators, and sociodemographic variables. Our findings demonstrate, firstly, that on-street population extracted from street view images outperforms other measures in assessing the risk population for street theft due to its superior ability to indicate outdoor activities. Moreover, this advantage of on-street population holds true during both daytime and nighttime hours. Secondly, differences in the impact of streetscape physical environment variables and control variables on street theft counts are observed across daytime and nighttime hours. On-street population proves to be a promising alternative for measuring the risk population of street crime victimization, and temporal variations should not be overlooked in future research endeavors.