In the daily operation of traffic, traffic control authorities may temporarily deploy mobile traffic lights at intersections due to power system maintenance, occasional power supply failures, or traffic control measures caused by road construction. These mobile traffic lights differ significantly in physical form from the fixed traffic light towers we are accustomed to seeing high in the air. They typically consist of a base, battery pack, support pole, and signal lights, with an overall height significantly lower than standard signal lights. These mobile traffic lights are placed directly on the ground, with the height of their illuminated units possibly only around waist level for a human or slightly higher than the engine hood of an average sedan. In such scenarios, human drivers can rely on experience to identify their presence, but for intelligent driving systems that heavily depend on preset rules, geometric models, and line-of-sight perception radius, this presents a significant challenge.

Occlusion Issues at the Physical Perception Layer and Sensor Fusion Mechanisms

The primary impact of mobile traffic lights on intelligent driving systems lies in their 'atypical' spatial positioning. According to mainstream specifications in the traffic equipment market, mobile solar signal lights have a wide range of adjustable installation heights. For example, the common TSU series devices can be adjusted from a collapsed height of around one meter to an extended height of over three meters. However, during emergency deployments triggered by power failures, in pursuit of installation efficiency, the equipment is generally set at a lower physical height. When an autonomous vehicle follows a large bus or van into an intersection, the visual camera, limited by its downward viewing angle and physical obstruction from the preceding vehicle, is highly likely to lose continuous visual tracking of these low-lying signal lights.

As the core sensor for identifying the color characteristics of traffic lights, the camera's detection effectiveness is highly dependent on 'line-of-sight accessibility.' If the preceding vehicle obstructs the key illuminated area of the signal light, even if this obstruction covers only about 30% of the signal light's illuminated surface, the autonomous driving system may fail to recognize it due to the inability to extract a complete circular or arrow outline. In complex urban scenarios, the visual distortion caused by large vehicle obstructions is not simply a matter of 'not seeing.' When the outline of a large vehicle fills the sensor's field of view, the autonomous driving system struggles to distinguish whether the object ahead is a static traffic facility or a trailer moving in tandem with the large vehicle. To address this issue, solutions proposed by companies like Didi, Huawei, and Tesla generally incorporate a weight allocation logic for multi-sensor fusion.

When the confidence level of the visual sensor declines due to occlusion, the autonomous driving system significantly increases the data weight of LiDAR and occupancy networks. Although LiDAR cannot directly 'see' the color of traffic lights, it can emit laser pulses millions of times per second to construct a three-dimensional point cloud model of the intersection ahead. Even if the traffic light is partially obstructed, LiDAR can still perceive the presence of a physical entity with specific geometric characteristics at the center or edge of the intersection. The existence of this physical entity triggers the occupancy network algorithm at the bottom layer (bottom layer, translated for clarity as 'core layer') of the intelligent driving system. The occupancy network divides three-dimensional space into countless tiny 'voxel' cubes, directly determining whether these cubes are occupied by physical objects. This logic bypasses the 'semantic recognition' trap, meaning the system does not need to immediately know what the low-lying object is; it only needs to know that there is an 'impassable' obstacle, thereby establishing the first line of defense against collisions at the perception level.

When dealing with perception voids caused by occlusion from preceding vehicles, some solutions also employ 'dynamic data augmentation' techniques. These techniques use algorithms to adjust the camera's light-sensing parameters in real-time, attempting to extract weak color feature signals from complex backlight or high-glare backgrounds. For example, when a backlight environment occupies a large portion of the lens's field of view, the system performs stepwise compensation for brightness and contrast in the local area suspected of containing traffic lights, ensuring that the color saturation of the illuminated units is sufficient to trigger the threshold of the classification neural network.

Handling Semantic Conflicts Between Map Prior Information and Real-Time Perception

For autonomous driving systems, the second core challenge posed by mobile traffic lights is the logical judgment of 'true and false signals.' In most vehicles with navigation-assisted driving functions enabled, the navigation map not only provides route guidance but also stores a large amount of static infrastructure information, known as 'map prior.' When a vehicle approaches an intersection, based on map data, the system has already been informed in advance of the presence of a set of fixed light poles at that location. However, after a power outage, these fixed light poles go dark, and a temporary traffic light not recorded on the map appears on the ground.

In such a scenario, a rigorous 'logical game' unfolds within the autonomous driving system. Traditional intelligent driving solutions may overlook real-time ground signals due to excessive reliance on map information, causing the vehicle to accelerate at a temporary red light. However, current autonomous driving systems, especially those adopting 'light map' or 'mapless' architectures, adhere to the core principle of 'real-time perception dominance, map auxiliary reference.' When the system detects that the signal lights marked on the map are invisible or extinguished, while the perception module captures a temporary traffic light on the ground, it automatically reduces the trust weight of map information.

This conflict resolution logic relies on Road Topology Reasoning (RCR) technology. The RCR network can, like humans, infer legal driving paths in real-time based on the geometric characteristics of the intersection, the intersection pattern of lane lines, and the position of the intersection stop line. Even if the position of the mobile traffic light on the ground deviates from the standard coordinates on the map, the RCR network can recognize its regulatory role in the current spatiotemporal context. For example, if this temporary traffic light is placed at the center circle of the intersection, the system will use a logical model trained through deep reinforcement learning to determine the semantic signal of the object, indicating that the current traffic flow is under temporary regulation.

To enhance the stability of this reasoning, autonomous driving systems employ a 'crowdsourced update' dynamic mechanism. If the first vehicle passing through the intersection detects a discrepancy between map information and actual road conditions (i.e., fixed lights are off, and a temporary light is on the ground), it uploads this discrepancy feature to a cloud server. The cloud-based large model rapidly processes these data from real traffic flows, generates a temporary 'live layer,' and distributes it to subsequent vehicles approaching the intersection. This mechanism enables subsequent vehicles to be informed in advance of the presence of mobile traffic lights at the intersection through the temporary patch downloaded from the cloud, even if their vision is completely obstructed.

The autonomous driving system's judgment of temporary traffic lights undergoes several weighing stages. The first stage is perception consistency checking. The system compares whether the illuminated signal captured by the camera coincides with the physical entity detected by LiDAR. If the camera sees a red light, but LiDAR indicates that the location is empty (possibly due to road reflections), the system filters out this false signal. The second stage is spatial logic verification. The intelligent driving system calculates whether the detected signal light height falls within a reasonable threshold. For ground-based mobile traffic lights, although their height is below the normal range, once identified as a traffic control facility, their priority is elevated to the highest level.

This shift towards 'perception dominance' means that autonomous driving systems are becoming increasingly similar to experienced drivers, no longer mechanically executing map instructions but possessing the ability to adapt flexibly to on-site conditions. This flexibility is crucial for ensuring driving safety in chaotic intersections after power outages.

Spatiotemporal Sequence Memory and Socialized Intention Reasoning

Of course, in scenarios with mobile traffic lights, there is also an extreme case of object occlusion where the mobile traffic light is completely obstructed. In such cases, the intelligent driving system does not immediately enter a 'blind' state but instead introduces a memory mechanism based on spatiotemporal sequences. Simply put, the autonomous driving system does not rely solely on the current frame but also retrospectively considers perception results from the past few seconds or even tens of seconds.

If the vehicle saw the signal state of the mobile traffic light when it was 50 meters away from the intersection, then even if the view disappears when approaching the intersection to within 20 meters, the recurrent neural network (RNN) or Transformer model within the autonomous driving system will still retain the 'hidden state' of this object. This state memory includes not only the object's position but also its last observed color. As long as the time span falls within the safety range set by the algorithm, the autonomous driving system will maintain its existing understanding of the traffic rules at this intersection and continuously update the relative coordinates of the traffic light in virtual space based on the vehicle's driving distance and speed.

Another solution, 'socialized intention reasoning,' can also address this issue. In complex traffic flows, the signal state of traffic lights is reflected in the behavioral patterns of surrounding vehicles. When the autonomous driving system cannot see the traffic light, it begins to frequently observe the movements of preceding vehicles and the 'collective choices' of surrounding traffic. For instance, the autonomous driving system monitors whether the brake lights of preceding vehicles are frequently illuminated and whether they accelerate decisively or hesitate and move slowly when crossing the stop line.

Through this monitoring of other vehicles' behaviors, the autonomous driving system is essentially conducting indirect signal inference. If all lanes ahead are queuing, and lateral traffic flows begin to move, even if the vehicle cannot see the temporary traffic light placed on the ground at all, the system can infer that the current signal must be red. This decision-making logic based on 'social evidence' prompts the autonomous driving system to calculate an 'environmental consistency probability.' When observing preceding vehicles decelerating and the map indicating an intersection, even with missing visual perception, the system adopts a conservative defensive driving strategy.

This socialized intention reasoning algorithm pathway involves two core steps. The first is trajectory prediction analysis. The autonomous driving system not only tracks the current position of preceding vehicles but also predicts their motion trajectories for the next 3 to 5 seconds. If the prediction indicates that the preceding vehicle will come to a stop before the stop line, the system initiates a deceleration decision in advance. The second is multi-vehicle behavior aggregation. The system simultaneously analyzes the vehicle states in the left, middle, and right lanes. If vehicles in all three lanes exhibit stopping behavior, the certainty of the environmental signal significantly increases.

To achieve this goal, autonomous vehicles must possess powerful computational resources to process vast amounts of data in parallel. In Tesla's FSD system, the neural network not only learns to recognize objects but also to predict the content of the next frame of images. If the mobile traffic light that should have appeared in the next frame is obstructed by a large vehicle, the model generates a predictive copy internally using existing spatiotemporal weights. Although this 'mental filling' ability carries some risk of error, it is more reliable than a simple signal interruption when dealing with short-term occlusions.

Final Thoughts

The negative impact of mobile temporary traffic lights on autonomous driving systems is gradually being mitigated through collaborative evolution at the perception, map, and decision-making layers. Although these low-lying, easily obstructed devices are indeed a 'weak spot' for autonomous driving at this stage, with the maturation of technologies such as occupancy networks, spatiotemporal sequence memory, and socialized reasoning, intelligent driving vehicles will also acquire the ability to navigate safely in imperfect environments.

-- END --