AI News

Shanghai HiSilicon launches the CS6286 image sensor, reducing sentry mode standby power consumption to under 50mW

Semiconductor Industry ObserverUpdated Nov 13 15:59

AI summary

Shanghai HiSilicon has launched the lightweight intelligent image sensor CS6286, integrating AI inference inside the sensor, reducing sentry mode standby power consumption to under 50mW, and extending standby duration from 1 to 2 days with traditional solutions to more than 7 days. The chip supports 8-megapixel, 4K@30fps HDR recording, has a built-in AI engine with equivalent 50GOPS@int8 computing power, and adopts a two-level wake-up mechanism combining regional motion detection and sensor-side AI recognition. The company says the humanoid target detection accuracy is above 95%. During parking standby, the main control SoC is completely powered down, and the sensor continuously buffers the most recent 10 frames of VGA RAW images; after determining a valid target, it wakes the main control within 50 milliseconds and retains the images from before the trigger. The CS6286 has entered mass production, and the accompanying AI training kit supports customers in adding up to 3 custom target classifications.

Why read this

AI score 51

The original text provides the specific parameters of the sensing-computing-storage integrated architecture and the two-level wake-up mechanism, allowing readers to judge how on-device AI can be implemented in low-power standby scenarios.

Topics

Tags

Body · AI translation

HiSilicon's all-new CS6286 sensor enables in-car sentry mode with "one charge per week"

Source: Bandaozongheng Published: 2026-11-13 14:27

Sensor Electric Vehicle Generate Poster

With its "sensing-computing-storage integrated" architecture and AI precise wake-up technology, the HiSilicon CS6286 reduces the standby power consumption of the dashcam sentry mode to under 50mW, successfully breaking through the battery life bottleneck and achieving long-cycle monitoring of more than 7 days.

Users who have used a dashcam's "parking monitoring" or "sentry mode" have almost all encountered this dilemma: with the feature on, the vehicle triggers a low-battery warning after being parked for two days; with the feature off, if the vehicle gets scratched, the culprit cannot be found.

This pain point that has long plagued the industry has never been rooted in the dashcam's imaging lens, but rather in the excessively high power consumption of the device's main control chip. In traditional solutions, to keep the lens on watch, the main control must remain in a semi-awakened state for long periods. Even when irrelevant movements appear in the frame—such as falling leaves, shifting light and shadow, or a stray cat passing by—the main control immediately starts up at full speed, rapidly draining the battery and generating large amounts of useless footage.

Recently, Shanghai HiSilicon launched a light intelligent image sensor, the CS6286, which supports 8-megapixel imaging and is primarily aimed at solving the low-power monitoring challenge of the dashcam "sentry mode." The biggest highlight of this chip is not the pixel count itself, but that it puts AI inference directly inside the sensor, creating a "sensing-computing-storage integrated" architecture. Put plainly, the sensor itself can decide whether to record and whether to wake the main control, without bothering the main chip every time.

First, the fundamentals: 1/1.8-inch, starting from 4K HDR

The CS6286's physical resolution is 3840×2160, a standard 4K specification. In HDR mode it supports 4K@30fps recording, and at full frame rate without HDR it can reach up to 8M@60fps. The sensor target size is 1/1.8 inch, with a single pixel size of 2.0μm.

What level is this specification among automotive sensors? Many mainstream aftermarket dashcam sensors still use 1/2.8-inch or even 1/3-inch targets. With a target one size larger, the amount of light entering naturally increases, and in low-light scenes such as dusk and underground garages, there will be far less noise. Moreover, for the AI analysis to be performed later, the higher the signal-to-noise ratio of the original image, the lower the misjudgment rate—this is a chain reaction. CS6286 imaging effect (Source: Shanghai HiSilicon)

Sentry mode: from "lasting two days" to "lasting a week"

The so-called dashcam sentry mode essentially means that after the car is turned off, it can still keep an eye on the surroundings of the vehicle and automatically record if someone touches or scratches the car. But this feature has a long-standing problem: it consumes too much power.

In traditional solutions, sentry mode relies on the main chip staying awake, or on the sensor performing coarse-grained motion detection. Power consumption cannot be reduced, and in practice it can only last 24 to 48 hours. If the car is parked for three or four days before being driven again, the battery may be drained, and many car owners simply turn this feature off in actual use.

CS6286's core design approach is to take over the core step of "movement judgment" from the external main control chip and complete the processing locally at the image sensor end. Inside this chip is an integrated lightweight AI engine with equivalent computing power of 50GOPS@int8, specifically responsible for completing real-time image analysis directly on the sensing side. The specific operating logic of this solution is divided into a two-level trigger mechanism:

The first level is first-level wake-up, which uses hardware-level regional motion detection. The chip continuously performs background frame modeling in the background, shielding out invalid interference such as leaf swaying and light-shadow changes at extremely low power consumption. Only when real physical motion appears in the frame—such as a human figure passing by—is it marked. But this level only determines "whether something moved" and cannot tell what is moving, so a second level is needed to close the loop.

The second level is second-level wake-up: after the first-level wake-up gives a signal, the built-in sensing-side AI engine immediately intervenes to determine whether the movement just now was a person, a car, or purely light-shadow interference. Official data shows that in common scenarios such as parking lots, roadsides, and courtyards, the detection accuracy for human-shaped targets is above 95%, and the overall detection accuracy is likewise above 95%.

This means that the main chip is awakened only when it truly needs to wake up and work; ordinarily, the sensor itself keeps watch at low power consumption.

Pre-recording power saving: locking in key footage with minimum energy consumption

Wake-up capability alone is far from sufficient. The most fatal problem with sentry mode is being "half a beat slow": by the time the system reacts and starts recording, the person or vehicle involved has already left, and in the end only their back may be captured or the key footage may be completely lost.

CS6286 has optimized the entire chain process in a targeted way: during the parking monitoring stage, the main control SoC is completely powered down, and the sensor operates in a single-chip low-power mode, using built-in SRAM to circularly cache the most recent 10 frames of VGA-resolution (640×432) RAW images, while keeping the sensing-side AI engine continuously on watch. Once the AI determines that a valid evidence-collection target appears, the sensor wakes the main control through a dedicated wake-up channel within 50 milliseconds, while fully saving the 10 frames cached before the trigger for the main control to read, fundamentally avoiding the loss of key evidence footage.

The video recorded in this way starts from the first frame before the trigger, so there will be no missing "previously on" context.

With this whole combination of measures, standby power consumption can be reduced to under 50mW, and the official claim is "one charge per week." The battery life of sentry mode is directly extended from 1 to 2 days in traditional solutions to more than 7 days. For users, this means that after parking the car for a weekend trip out of town, the dashcam is still working normally when they return.

Other details and deployment status

For the remaining parameters, the CS6286 uses a WLCSP package, outputs 12-bit RGGB RAW data, and its interface is MIPI D-PHY, supporting 1/2/4/2×4-lane configurations, with a maximum of 2.5Gbps per lane; the shutter type is rolling shutter. For the main control of an aftermarket dashcam, this interface offers relatively friendly compatibility, and the 4-lane bandwidth is also sufficient for real-time 4K@30fps HDR output. Key parameter configuration of the CS6286 (Source: Shanghai HiSilicon)

Also worth mentioning is that Shanghai HiSilicon has provided the CS6286 with an AI training suite covering the entire workflow of model degradation optimization, retraining, and compilation, supporting customer-defined target categories, with up to 3 additional target classifications. That is to say, in addition to the default human and vehicle detection, customers can also add target categories they particularly care about according to their own scenario requirements. The entire toolchain is complete and easy to use; customers only need to input a custom dataset and corresponding parameters to directly output a model parameter configuration file that can be deployed directly.

At present, the CS6286 has entered the mass production stage, and the training suite delivered with the chip is also ready, so customers do not need to build models from scratch during introduction.

After all, competition in automotive sensors in recent years has long ceased to be merely about pixels and dynamic range; the industry has officially upgraded toward "edge intelligence." The sensor itself can complete image content recognition and judgment, which can free up the computing power of the main control chip to carry more functions, while achieving bidirectional optimization of power consumption performance and user experience. The new CS6286 is precisely a representative product sample that is among the first to run through this technical path.

This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our reproduction is solely for sharing and does not mean that we approve of or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly; as a neutral platform service provider, we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

Bypassing HBM and liquid cooling! European AI chip company launches Titania architecture 2026-11-13 15:21

Total investment of 46 trillion won, Samsung's new plant covers HBM4E and next-generation capacity needs 2026-11-13 15:18

Applied Materials FY2026 Q4 revenue of $10.41 billion, AI drives equipment demand to a new high 2026-11-13 15:17

Source

Semiconductor Industry Observer