Many researchers have delved into the matter of how police enforcement influences drivers’ driving speeds. Ref. [
11] investigated factors influencing speed choices on rural roads in Norway with an 80 km/h speed limit. It looked at how drivers’ perceptions of police enforcement, penalties for speeding, and other drivers’ speeds affect their choices. The study found that making most other drivers slow down or increasing law enforcement had the most significant impact on reducing individual speed choices, while stricter penalties had only a minor effect. Ref. [
12] examined the impact of police presence on speeding in urban areas using a realistic-looking police cut-out named “Constable Scarecrow” in British Columbia, Canada. The findings showed that deploying the cut-out along major roads helped reduce speeding among motorists. Ref. [
13] looked at how police saturation enforcement impacts speeding on a highway corridor in Western Canada. They used radar devices at different locations to measure speeds during enforcement and non-enforcement periods. The results showed that police saturation enforcement effectively reduced average vehicle speeds and the proportion of speeding vehicles in the enforcement area, contributing to discussions on policing and road safety. However, there is still a scarcity of relevant research on crowdsourced data for police enforcement.
2.1. Crowdsourced Data Usage in ITS
Crowdsourced data from social media apps such as Twitter have become an emerging data source to provide new support for predicting or managing traffic issues due to the rapid growth of smartphone users. An increasing number of research has investigated the application of such crowdsourced data in traffic management. Ref. [
14] used crowdsourced data from smartphone applications, GIS-based web interfaces, and weather sensors to model the individual mobility decision processes. The model is regarded as a potential platform for personalized travel management in smart cities, as well as a communication tool between cities and users. Ref. [
15] mined crowdsourced media data from Twitter and Foursquare to look into the spatial and temporal patterns of human activities in a city, and showed the importance and usefulness of crowdsourced data in analyzing people’s activities. Ref. [
16] proposed that mining social media data can be a basic low-cost supplement and convenient solution for ITS. More than 1 million tweets were collected over 3 months, and an Arabic Twitter content analysis framework was proposed to tackle the problem of missing the location information of traffic-related incidents in the tweets. Ref. [
17] developed a machine learning method to predict traffic evolution after accidents based on the user-generated crowdsourced data (UGCD) provided by navigation apps that have interfaced for users to report traffic incidents, and showed the efficiency of using UGCD for the real-time analysis of traffic accidents. Ref. [
18] assessed the speed data based on a crowdsourced navigation system, Waze, and conducted a case study in Sevierville. They showed that the posted speed on Waze is a good representation of actual speed. Ref. [
19] proposed an innovative machine learning framework to extract traffic-related information from social network crowdsourced data for traffic incident detection. These studies and others [
20,
21,
22] indicate that crowdsourced data from social media or smartphone applications are one of the promising data sources for managing smart cities and transportation systems.
2.2. Traffic Speed Prediction
Accurate traffic speed prediction is an important component in ITS, as it offers useful information to reduce traffic congestion by providing route guidance to travelers [
23]. Extensive research has been conducted on using available datasets to predict traffic evolution. The difficulty of predicting traffic speed in urban road networks lies in (i) accurately extracting temporal and spatial features of traffic networks, and (ii) adequately considering the impact of external factors on traffic flow from multiple sources of data, such as weather, social events, accidents, etc.
Traditionally, researchers use mathematical statistics to analyze and predict traffic states, such as the ARIMA model [
24,
25], the Kalman filter algorithm [
26], the hidden Markov model [
27], the Bayesian network [
28], etc. These statistical techniques could model the traffic conditions using relatively small-scale datasets but have limited ability to capture the nonlinear characteristics of traffic data.
With the development of data collection and computing power, most recent works have focused on data-driven models, particularly from traditional machine learning models [
29,
30] to deep learning models [
31,
32] that could perform the prediction task well based on historical traffic databases. The advantages of machine learning models are the ability to handle multidimensional data, implementation flexibility, generality, and strong predictive capabilities [
33]. However, they cannot capture the spatial correlations of road networks well. In addition, compared with deep learning methods, the prediction accuracy of machine learning is relatively lower due to the shallow structure.
Recently, more advanced and powerful deep learning models have been applied to traffic prediction. Ref. [
34] represented large-scale traffic networks as images, and adopted the deep learning architecture of convolutional neural network (CNN) to extract the spatio-temporal traffic features contained in the images. The traditional CNN can use a fixed-size learnable convolution kernel, which effectively describes the spatial characteristics of Euclidean data, such as text, sound, and images, and extracts useful information from them [
35]. However, real road networks are difficult to meet the approximate grid shape in space, and a large amount of traffic data is complex non-Euclidean data [
36].
In order to incorporate the topology of road networks and exploit graph structure information, the graph convolutional neural network (GCN) is extended and applied to traffic prediction [
37,
38]. The principles of GCN are to regard the transportation network as a graph and recognize the connectivity of roads by an adjacency matrix, then extend the convolution operation on the graph structure to aggregate the information of each node, which is proved to be more effective than a grid-type convolution on capturing topology features of transportation networks and forecasting traffic speed [
39]. In order to take into account the temporal characteristics of traffic data simultaneously, the gated recurrent unit (GRU) is also widely used in traffic speed prediction combined with GCN [
31]. GRU is an improved version of the recurrent neural network (RNN), which could process sequence data and can reduce the vanishing gradient problem, while preserving long-term sequence information [
40]. The efficiency of GCN and GRU as traffic prediction models has widely been proved using real traffic data [
39,
41].
In addition, traffic forecasting is more challenging than other spatiotemporal forecasting problems because it involves many external factors, which affect traffic states. Ref. [
42] constructed the traffic speed prediction model considering the day of the week and POI. Ref. [
43] proposed a model with bidirectional long short-term memory (LSTM) and a complex attention mechanism to predict the urban traffic volume, combined with weather conditions and event information as external features to further improve the prediction precision. A traffic graph convolution operator was proposed in [
44] in order to extract the local features and combine the physical features of the road network. Ref. [
45] modeled external factors as dynamic and static attributes, and designed an attribute augmentation unit to encode and integrate these factors into a spatio-temporal graph convolution model, and demonstrated the effectiveness of considering external information in the traffic speed prediction task. However, due to the complexity of built environment, there are various types of latent factors which affect the driver behavior, thereby affecting the traffic speed and adding uncertainty to traffic prediction problems.
To summarize, works in the existing literature do not fully consider the influence of external factors on traffic. In particular, how the crowdsourced data about police speed checks report in navigation apps influence drivers’ driving behavior and overall traffic speed on the road network has not been investigated yet.