Sensor fusion is the practice of combining readings from two or more sensors into a single estimate of the world that is more reliable than any individual sensor. An autonomous vehicle carries cameras, radar, often LiDAR, plus inertial and wheel-odometry sensors, and each has different failure modes: cameras struggle in glare and darkness, radar has coarse resolution, LiDAR degrades in heavy rain. Fusion is how a perception stack gets a clean picture out of imperfect, disagreeing inputs. In the patent record, fusion inventions are classified in the conjoint vehicle-control subclass B60W — whose CPC title is "CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION" — and in image- and signal-processing perception classes, depending on whether the claim emphasizes the vehicle action or the data processing.

As with most autonomy concepts, the claims rarely use the marketing phrase. "Sensor fusion" describes an objective; the claim describes a mechanism. The mechanism almost always opens with receiving data from a first sensor and a second sensor, then doing something that joins them. One especially clean fusion sub-pattern is cross-validation: rather than blending two readings into an average, the system uses one sensor to check another and flag a fault. US9555740B1, "Cross-validating sensors of an autonomous vehicle," a granted patent assigned to Google LLC, claims this directly.

receive first sensor information from a first sensor configured to detect objects in a vehicle's environment; receive second sensor information from a second sensor configured to detect objects in a vehicle's environment; identify a first object and a first characteristic value of the first object using the first sensor information; identify the first object and a second characteristic value of the first object using the second sensor information— US9555740B1, claim 1, source

The rest of claim 1 completes the cross-validation loop. Having extracted a characteristic value for the same object from each of the two sensors — "wherein the first characteristic and the second characteristic are of a same type" — the system computes "a deviation value" between them and compares it against "a deviation threshold value identifying an acceptable amount of difference." When the deviation exceeds the threshold, the claim concludes, the system determines "that the first sensor has had an operational failure." That is fusion used not to build a richer percept but to police sensor health: two sensors looking at the same object should agree within tolerance, and disagreement beyond tolerance is itself information. The claim fences off a specific safety mechanism — sensor self-diagnosis through cross-comparison — rather than fusion in the abstract.

Two flavors of fusion in the claims

Reading fusion patents, you encounter two broad claim shapes. The first is combinative fusion: the claim merges multiple sensor streams into one enriched representation — a denser point cloud, a tighter object bounding box, a fused track. The emphasis is on producing a better estimate. The second is comparative fusion, the cross-validation pattern above, where the claim keeps the streams separate and uses their agreement or disagreement to decide something — whether a sensor failed, whether an object is real, whether to trust a reading. US9555740B1 is squarely comparative: it never blends the two characteristic values, it differences them. Knowing which shape a claim takes tells you what it actually covers. A comparative cross-validation claim does not read on a combinative fusion system, and vice versa, even though both would be described in a press release as "sensor fusion."

Why fusion claims are a crowded field

Sensor fusion is old. The mathematical core — Kalman filtering, Bayesian estimation, track-to-track association — predates autonomous cars by decades and runs through aerospace and defense literature from the mid-twentieth century onward. That deep prior art means a modern fusion patent generally cannot claim "combining sensor data" broadly; it must claim a specific combination, a specific decision rule, or a specific application. US9555740B1's narrowing move is the deviation-threshold fault-detection logic; another patent's might be a particular sensor pairing, a confidence-weighting scheme, or a specific edge case. For an IP reader, the practical consequence is that the B60W and perception fusion landscape is dense with narrow, implementation-specific claims rather than broad foundational ones, and the competitive question is rarely "who owns sensor fusion" — no one does — but "who owns the specific fusion mechanism this product depends on." The answer is always in the limitations of the independent claim: which sensors, what comparison, what threshold, what action. Strip those and there is no claim left.

The threshold itself is a place where fusion claims quietly reveal the engineering. US9555740B1 does not just compare two readings; it compares the deviation against "a deviation threshold value identifying an acceptable amount of difference between the first and second characteristic values." That single phrase encodes a design decision: sensors never agree perfectly, so the system must define how much disagreement is normal before it treats a discrepancy as a failure. Set the threshold too tight and the system cries wolf on healthy sensors; set it too loose and a genuinely failing sensor slips through. The claim fences off the use of such a threshold for fault determination, and the dependent claims of patents in this family frequently narrow further to specific threshold-setting or specific characteristic types. For an IP reader, the threshold limitation is a useful tell: a comparative-fusion claim that recites a tolerance is claiming a diagnostic mechanism, not a perception-enrichment mechanism, and the two cover different products.

One last reading note. Because fusion sits at the boundary between perception and control, the same invention can be classified under B60W (the vehicle-control angle) or under an image- or data-processing class (the computation angle), and many fusion patents carry symbols in both. When mapping a fusion portfolio, search across both the vehicle-control and perception classes rather than relying on a single class, or you will miss half the estate.