Investigating the effect of people on the street and streetscape physical environment on the location choice of street theft crime offenders using street view images and a discrete spatial choice model

As street crime occurs in the street, it is reasonable to assume that the location choice of street crime offenders is affected by streetscape conditions and people on street. However, this issue has not been investigated by previous research, possibly because fine-grained streetscape data are hard to obtain. Traditional data-gathering methods like questionnaires, field surveys, and manual audits are inefficient and only applied to small areas. Social media and mobile phone data have recently been used to measure the ambient population. However, they are unable to distinguish between indoor and on-street people. To overcome these limitations, the present research applied an integrative deep learning algorithm combining an object detection network and a semantic segmentation network to extract on-street people and physical environment elements from fine-grained street view images (SVIs) in a large Chinese city. The extracted elements include fences, walls, windows, grass, sidewalk, and plants. Controlling the influence of residence-crime proximity, crime attractors, generators, detractors, and socioeconomic features, we constructed a discrete spatial choice model to investigate the influence of people on the street and the streetscape's physical environments on the location choice of street theft crime offenders. Results reveal an improvement in model performance after the streetscape variables are considered. Therefore, the streetscape context is essential for understanding offenders' preferences for crime locations. Specifically, the number of people on the street presents a significantly positive relationship with the offenders' preferences. Fences and plants have significant and positive effects on attracting criminals. Grasses and sidewalks negatively affect offenders' location choices. Walls and windows do not significantly affect criminals' crime location choices. Additionally, the associations between most control variables and offenders' preferences for crime locations conform to previous research findings. As the first attempt in combining SVIs, deep learning algorithms, and discrete spatial choice model, this study makes a contribution to the extant crime location choice literature.