The way in which the occurrence of urban traffic collisions can be conveniently and precisely predicted plays an important role in traffic safety management, which can help ensure urban sustainability. Point of interest (POI) and nighttime light (NTL) data have always been used for characterizing human activities and built environments. By using a district of Shanghai as the study area, this research employed the two types of urban sensing data to map vehicle-pedestrian and vehicle-vehicle collision risks at the micro-level by road type with random forest regression (RFR) models. First, the Network Kernel Density Estimation (NKDE) algorithm was used to generate the traffic collision density surface. Next, by establishing a set of RFR models, the observed density surface was modeled with POI and NTL variables, based on different road types and periods of the day. Finally, the accuracy of the models and the predicted outcomes were analyzed. The results show that the two datasets have great potential for mapping vehicle-pedestrian and vehicle-vehicle collision risks, but they should be carefully utilized for different types of roads and collision types. First, POI and NTL data are not applicable to the modeling of traffic collisions that happen on expressways. Second, the two types of sensing data are quite suitable for estimating the occurrence of traffic collisions on arterial and secondary trunk roads. Third, while the two datasets are capable of predicting vehicle-pedestrian collision risks on branch roads, their ability to predict vehicle safety on branch roads is limited.