Adaptive traffic control is one of the few urban technologies with a genuinely well-evidenced record. Deployed on the right network it reduces delay, smooths flow and improves journey time reliability, and it does so at a fraction of the cost of new road capacity. It is also, in the wrong conditions, an expensive way of measuring a problem you already knew you had.
Dhaka contains both conditions, sometimes within the same corridor. That makes the interesting question not whether the technology works, but where it works and what has to be true first.
What adaptive signalling actually does
A modern adaptive system observes demand at each approach to a junction, predicts arrivals, and reallocates green time continuously across a coordinated group of junctions. Its gains come from three sources: eliminating green time served to an empty approach, holding progression along a corridor so platoons of vehicles are not repeatedly stopped, and responding to incidents faster than a fixed plan can.
Every one of those gains depends on the same precondition. The system must be able to allocate right of way and have that allocation obeyed. It is a scheduling technology, and a schedule that nobody follows is not a schedule.
The enforcement precondition
Where the binding constraint on a corridor is not the timing plan but the behaviour at the junction — blocking the box, opposed movements, informal stopping across the mouth of the intersection, pedestrian crossing at will along the whole link — adaptive control cannot recover the capacity it is designed to recover. The signal releases a movement that is already obstructed. The optimisation is correct and the outcome is unchanged.
This is not a criticism of the city. It is an observation about ordering. The enforcement layer — automated detection, a functioning notice and penalty process, and a legal basis that survives challenge — is a precondition for the control layer to deliver, not an enhancement of it.
It is also the harder part to procure, because it is institutional rather than technical. Cameras are straightforward. A penalty process that is reliably issued, reliably served and reliably collected is a cross-agency reform with a data protection dimension attached.
Underneath that sits an unglamorous dependency that decides more automated enforcement programmes than any algorithm: the vehicle registry. Automated detection is only as good as the record it resolves to. Plates must be standardised and readable, the registry must be current, ownership transfers must be reflected in it, and the address against which a notice is served must be one at which the notice arrives. Where any of those fails, the system generates detections that cannot be converted into consequences, and behaviour does not change.
Heterogeneous traffic breaks the assumptions
Most adaptive algorithms were developed for lane-disciplined traffic composed of vehicles with broadly comparable acceleration and footprint. Dhaka’s traffic is none of those things. Non-motorised vehicles, three-wheelers, buses, motorcycles filtering continuously and pedestrians crossing mid-block share the same space without lane separation.
Two consequences follow. Detection is harder: loop-based counting assumes vehicles occupy discrete lanes, so vision-based classification is effectively mandatory and has to be trained on local traffic rather than imported. And saturation flow — the throughput a green phase can actually deliver — is far more variable than the models assume, which means parameters cannot be lifted from a reference deployment elsewhere.
This is where machine learning genuinely earns its place. Learning saturation behaviour from local observation, rather than assuming it, is a real advantage over conventional adaptive control.
Where the technology earns its cost
On corridors with physical separation, controlled access and enforceable discipline — expressway approaches, newer arterials, bus priority corridors, port and airport access roads — adaptive control delivers what it claims. These are also the corridors where economic value per minute saved is highest, because they carry freight and scheduled services rather than only discretionary trips.
The second sound investment is the layer beneath the optimisation: a command centre, a consistent detection estate and a data platform that multiple agencies can use. Even where signal optimisation is constrained, reliable measurement of flow, incident detection and journey time is valuable in its own right and is the evidence base for every subsequent intervention.
A sequencing view
Our reading is straightforward. Build the detection and data layer first, because it is useful under every scenario. Establish the enforcement process in parallel, because it is the slowest institutional item. Deploy adaptive control initially where discipline already exists and the gains are attributable. Extend it as enforcement extends.
One procurement point follows from this. A programme assembled in stages only works if the stages can be built by different parties over several years, which means open interfaces, documented data formats and ownership of the data resting with the city rather than the supplier. A closed system delivers the first phase efficiently and prices every subsequent one.
A programme built in that order produces measurable results at each stage. A programme that installs the optimisation layer first produces a very well-instrumented account of congestion, and a difficult conversation about why it has not changed.