By Transport Department, The Hong Kong University of Science and Technology, and the builder, QTC Traffic Technology Limited and Logistics and Supply Chain MultiTech R&D Centre

Background

In 1977, Transport Department (TD) commissioned the GEC 4080 digital computer to manage eighty (80) nos. of signalised junctions in the West Kowloon District, which was the first area traffic control (ATC) system commissioned in South East Asia. Since then, on-street traffic signals have been progressively coordinated to maximise traffic flow, and traffic signal timings can be adjusted at the control centre to suit varying traffic needs. The ATC system also provides round-the-clock equipment monitoring to allow expeditious rectification of traffic signal faults.

In the 1980s, with a view to enhancing the efficiency of traffic signals operation, Vehicle Actuated (VA) control, employing inductive loop detectors and pedestrian pushbuttons, were installed to effectively reduce the unused vehicular and pedestrian green times. VA was also deployed at shared junctions with Light Rail Transit (LRT) in Tuen Mun & Yuen Long (TM&YL) Districts, providing priority to LRT while serving other vehicular and pedestrian without approaches. To maintain coordination whilst benefiting from VA, TD developed a semi- VA algorithm, which made use of real-time detector data to extend or cut stage times early, while keeping the cycle length unchanged. In the late 80s, TD installed the first traffic adaptive control (TAC) SCOOT¹ ATC system on Hong Kong Island, deploying TAC at junctions in Causeway Bay and Central. TAC SCOOT is more versatile than semi-VA in that it could optimise split, cycle and offset timings for groups of signals based on real time traffic conditions. Since the 1990s, another TAC algorithm SCATS² has been implemented in Kowloon Districts, and gradually expanded to New Towns including Tsuen Wan, Sha Tin, Tai Po & North and Tseung Kwan O Districts. TM&YL Districts have also implemented SCOOT since 2008. Nonetheless, TAC or VA has not been prevalent across the territories due to disruption during installation and subsequent maintenance of the inductive loop detectors.

Past Trials of Off-ground Detectors

Noting the drawback of loop detectors on traffic control, TD continued to look for replacement detection technologies for traffic signals control. In 2000, TD trialled a video detection system at Peak Road / Plunkett’s Road junction, but was unable to overcome the challenging junction layout, vehicle headlights, and misty weather conditions at the Peak. It eventually concluded that the prevailing video detection technology was immature for traffic signal control.

More practical off-ground detectors have emerged in the industry since 2000, including radar, magnetometer, video, thermal sensors, LIDAR. TD began deploying video detection in 2010 at various sites, including Hoi Bun Road/Lai Yip Street in the context of walkability for reducing the waiting time of pedestrians. Also, video installation was deployed at Link Road/ Broadwood Road junction to automatically place or cancel a demand for pedestrian green time with due regard to whether a pedestrian was detected in the waiting zone, so as to enhance the junction efficiency.

In 2018, video cameras were further deployed at both ends of Tai Tam Road (Dam Section) and associated upstream locations to track and count vehicles passing or queuing for making real-time adjustment of the vehicular green time allocated to the ‘one-lane two-way’ traffic on the dam. This smart traffic control system effectively reduced the traffic queues and overall delays compared with those observed under traditional traffic signals, making traffic signal control a viable solution to the long-standing traffic problem at that narrow road section.

Latest RTATSS trial in Tung Chung

In line with the Government smart city initiatives, TD has been conducting pilot projects and trials for real-time adaptive traffic signal control (RTATSS) in Hong Kong. In 2021, TD deployed off-ground detectors at five isolated junctions to assess the practicability and benefits of RTATSS, and found that 5-10% reduction of vehicular and pedestrian delays were achieved as compared with operations with no optimisation.

Taking one step forward, to ascertain that RTATSS is also applicable at closely spaced junctions which are common in Hong Kong, eight signalised junctions in Tung Chung operating on the SCATS platform, were fully equipped with off-ground vehicular and pedestrian detectors for traffic signals control. It was found that a reduction in vehicular delay of 5-11% was achieved compared with conventional fixed-time operations.

To facilitate the growth of smart junctions and utilising TD’s existing Area Traffic Control Centre (ATCC) and ATC systems facilities to monitor/control the smart junctions, the eight signalised junctions in Tung Chung were also designed as a test platform for:

1. deploying the existing SCATS operation interface to oversee also smart junctions;
2. opening up the traffic control platform allowing other (non-SCATS) traffic control algorithms to demonstrate their applicability in Hong Kong.

Eventually, Both (i) & (ii) were proved to operate satisfactorily.

The Tung Chung RTATSS provides enhanced traffic signal control via machine vision, Artificial Intelligence (AI) technology and alternative traffic signals control algorithms. Vehicle and pedestrian images collected by thermal/video detectors and processed by AI render vehicular flow rates and pedestrian density in waiting zones. Through a newly developed signal control algorithm, the Tung Chung RTATSS has utilised the computed traffic data to adjust stage times of junctions and optimised the signal operations.

Tung Chung RTATSS Architecture

The system comprises field detection/control equipment, and central system equipment described below:

Field equipment-

• Thermal and optical traffic sensors
• Edge computer for image processing - AI Unit (AIU)
• Edge computer for logical processing - Local Data Processor (LDP)
• Traffic Signal Controller (TSC) Input/Output interfacing module
• Communication network equipment

Central equipment-

• A Central Control Processor (CCP)
• Interface between RTATSS and SCATS
• Web Portal Server and user terminal
• System alarm reporting system

Tung Chung RTATSS overall System Architecture

Vehicle and Pedestrian Sensors

Video-based thermal-type and optical-type vehicle sensors are used to detect approaching vehicles at a distance around 50m from the stop-line. Thermal-type pedestrian sensors are also installed adjacent to pedestrian waiting zones at junctions to collect pedestrian flow data within the detection area. The use of thermal sensors provides greater privacy, as vehicle number plates and human faces are not captured. Both vehicle and pedestrian sensors are mounted on an existing traffic signal pole with an add-on extension pole reaching up to 6 metres above ground level. Both thermal and optical video signals are sent to the AIU for processing.

Artificial Intelligence Unit and Sensor I/O Unit

Artificial Intelligence Unit (AIU), upon collecting video images from sensors and performing image processing, produces structured traffic data for further analysis. Using its AI algorithm, the system is capable of classifying double-decker buses, coaches, lorries, private cars and motorcycles. It also records each object's entrance and exit time, and calculates the waiting duration. For pedestrian sensing data, it identifies pedestrians from the video stream to distinguish whether they are waiting to cross or passing by the waiting zone. It also calculates zone occupancy to provide data for traffic signal control.

Thermal Vehicle Sensor Image	Optical Vehicle Sensor Image	Thermal Pedestrian Sensor Image

Local Data Processor (LDP)

LDP controls the extension of current stage and issues demands for a pedestrian stage by examining the traffic densities of the target vehicle and pedestrian zones, and queue lengths of all approaches. These parameters are described as follows:

Vehicle Density (vDen)

vDen = vOcc / vMax

where vMax is the maximum vehicle density of the detection zone.

n is the number of vehicles of type k, where k is vehicle type 1 to 5.
pcu(k) is the pcu value of type k

Pedestrian Density (pDen)

pDen = pOcc / pMax
where pMax is the maximum pedestrian density of the pedestrian waiting zone and pOcc is the number of pedestrians waiting in the waiting zone.

a. Stage Extension Control
Stage extension control is based on traffic density of the Detection Area and Counting Area of associated traffic arms-

Detection Area
Detection Area is the maximum detection zone of the imagesensing equipment and is used to estimate traffic density and queue lengths.

Counting Area
The Counting Area is a subset of the Detection Area and is used to detect the ‘last’ vehicle of the current movement or stage. The Counting Area varies in according to the queue lengths of all conflicting demands, i.e. the higher the conflicting demand, the smaller the Counting Area.

Figure 1 Vehicle Detection & Counting Areas

Queue Length Detection
Vehicles queueing during the red phase are measured in terms of the vehicle density of the Detection Area. During the vehicle red phase, the AIU continuously detects the build-up of the queue within the Detection Area and hence the vehicle density, which is used to adjust the Counting Area size of the approaches displaying green.

Overall, the LDP provides far greater flexibility in programming and fine-tuning its traffic signal control algorithms compared with conventional VA control logic.

Central Control Processor (CCP)

The CCP is the core component of the system to carry out traffic plan calculations and execution via its algorithm. The CCP algorithm collects queue length and traffic flow data from associated LDPs, and derives an optimum cycle length, junction grouping (subsystem) plan, coordination (link) plans, and the need of executing an “action list” to support special control functions that suit traffic conditions.

The CCP also monitors the operational status of all system equipment and field equipment for alarm reporting by sending alarm information to SCATS via the SCATS ITS port. The alarm reporting subsystem integrates the Tung Chung RTATSS alarm reporting subsystem to SCATS, enabling all relevant field equipment and system alarm to be displayed in ScatsAccess, the graphical user interface of SCATS.

Third party Algorithm Integration

The Tung Chung RTATSS allows third party systems to retrieve traffic data and reconstruct traffic plans using their own traffic control algorithm. A sandbox equipment is provided for a third party system to test their algorithm before deployment on the CCP. Once testing is successfully completed, the third-party algorithm is loaded into the CCP and allows it to control live traffic signals.

HKUST’s (third-party) Algorithms in Tung Chung RTATSS

The real-time traffic data collected from multiple thermal and optical sensors at the eight junctions in Tung Chung has facilitated the development of advanced adaptive traffic control systems on an area-wide scale. The Hong Kong University of Science and Technology (HKUST) has developed recent research advancements in traffic signal control aimed at enhancing traffic performance at real-world sites, including the Tung Chung RTATSS, in collaboration with TD.

Figure 2 DISCO model - Tung Chung area

Figure 3 VISSIM model – Tung Chung area

Two Layers Control Algorithms
In Tung Chung RTATSS, HKUST has developed a two-layer control algorithm. At the lower layer, two comprehensive traffic models have been used: the self-developed Dynamic Intersection Signal Control Optimisation (DISCO) model and the VISSIM micro-simulation model. Concurrently, at the upper layer, an AI- and machine-learning-based approach for adaptive plan optimisation has been established.

DISCO is founded on a dynamic traffic model that accurately captures the three fundamental characteristics of traffic flow: spatial, temporal, and stochastic properties. Using DISCO, signal timing plan solutions have been developed for largescale operations, covering eight junctions in Tung Chung, employing enhanced algorithms based on AI and machine learning, supported by parallel computing to substantially improve operational efficiency.

The VISSIM model replicates traffic conditions in Tung Chung, incorporating elements such as public transport routes and stops, as well as driving and lane-changing behaviours. VISSIM was utilized to validate the performance of adaptive control plans derived from DISCO through API linking before on-site implementation.

Adaptive signal control was successfully achieved across various periods of the day. Following optimisation and validation, adaptive timing plans were integrated into the SCATS platform and the CCP for implementation. Appropriate number of timing plans, with different green times, cycle times and offsets, were encoded within the SCATS. In the CCP, for each cycle, traffic from multiple entries was classified into distinct traffic patterns, and optimal adaptive signal plans were triggered on a cycle-by-cycle basis for all eight junctions, thereby achieving coordination among them.

Evaluation of the benefits of RTATSS

HKUST has established a calculation method that estimates average vehicular delay using the queuing delay formula, utilising available detector information and signal timing data from the CCP.

The existing thermal sensors capture the waiting time of vehicles within the detection zone, which typically encompasses 50 meters beyond the signal stop-line. Vehicles in the detection zone are visible and their arrival times can be captured by the sensors (referred to as captured vehicles), while those queueing beyond the detection zone are not visible and their arrival times cannot be captured by the sensors (classified as uncaptured vehicles). The estimation of waiting time is accurate when traffic volume is low and all vehicles are detected. However, when queues extend beyond the detection zone, the waiting time of uncaptured vehicles may not be recorded, as they remain out of sight.

Utilising D/D/1 queue theory, HKUST used sensor information at the stop-line to determine the discharge rate and queue length, corresponding to the departure curve information. The arrival times of uncaptured vehicles were estimated based on uniform traffic arrival conditions, corresponding to the arrival curve information. By integrating the departure and arrival curve data, HKUST constructed the queuing diagram and estimated the average vehicular delay for all captured and uncaptured vehicles.

The maximum reduction in delays rendered by HKUST’s two layers control algorithms were found to be approximately -11% and -14% during peak and off-peak hours respectively.

Concluding remarks and the way forward

Given the experience gained and benefits demonstrated by RTATSS and third party algorithms, TD has scheduled to implement RTATSS at 50 isolated junctions. Further trials will continue at other busy junctions allowing more data to be collected to inform long-term strategies at existing junctions. Nonetheless, suitable new junctions in the New Development Areas (NDAs) will allow adequate provisions for RTATSS operations in one go as far as practicable.

¹ Split Cycle Offset Optimisation Technique (SCOOT) is a proprietary traffic adaptive control system developed by the UK’s Transport Research Laboratory.
² Sydney Coordinated Adaptive Traffic System (SCATS) is a proprietary traffic adaptive control system developed by the Transport for New South Wales Government in Australia.