Autonomous vehicles and ADAS

Automobile Engineering (B.E./B.Tech) · Electric and future vehicles · free study notes for exams, placements and interviews

In short

An autonomous vehicle senses its surroundings and drives itself with little or no human input, while ADAS (Advanced Driver Assistance Systems) are the electronic aids - adaptive cruise control, lane keep assist, automatic emergency braking - that assist a driver who stays responsible. The two sit on one continuum: SAE Levels 0-2 are assistance, Levels 3-5 are automation. The subject matters because the sensing, computing and drive-by-wire hardware it needs is now standard on mainstream Indian cars (Mahindra XUV700, Tata Harrier/Safari, Honda City e:HEV, MG Astor, Hyundai Creta) and because Bharat NCAP from October 2023 awards points for ADAS fitment. It is examined in automobile/mechatronics electives, GATE ME and IN control and instrumentation sections, and asked by Bosch, Continental, ZF, KPIT, Tata Elxsi, L&T Technology Services, Mercedes-Benz R&D India and Qualcomm in placement rounds.

Explained

Levels of driving automation (SAE J3016)

Level 0 - No automation: warnings only, no sustained control. Forward collision warning, blind spot warning, lane departure warning, parking sensors. Momentary intervention like AEB and ESC also counts as Level 0.
Level 1 - Driver assistance: automation of one axis - either longitudinal (adaptive cruise control) or lateral (lane centering), not both.
Level 2 - Partial automation: longitudinal and lateral together (ACC + lane centering = "highway assist"). The driver must supervise continuously and remains the legal driver. Almost every ADAS car sold in India today is Level 2.
Level 3 - Conditional automation: the system drives within a defined ODD and the driver may disengage, but must take over when the system issues a request to intervene. Legally the key break - the system, not the human, monitors the environment.
Level 4 - High automation: no human fallback needed inside the ODD; the system reaches a minimal risk condition (safe stop) by itself. Robotaxis in geofenced areas.
Level 5 - Full automation: no ODD restriction; steering wheel optional.
ODD (Operational Design Domain): the stated conditions - road type, geography, speed range, weather, lighting, traffic - in which the feature is designed to work. Outside its ODD the feature must not engage or must hand back.
DDT (Dynamic Driving Task) = lateral control + longitudinal control + object and event detection and response (OEDR). DDT fallback = what happens when it fails.
Levels describe a feature, not a vehicle: one car may have a Level 2 highway feature and a Level 4 parking feature.

Sensors: what each one measures and where it fails

Camera (monocular/stereo): passive, high resolution, the only sensor that reads colour, text and texture - lane markings, traffic signs, signal colour, vehicle class. Weak in glare, fog, night, heavy rain; monocular depth needs scale assumptions.
Radar (24 GHz short range, 76-81 GHz long range): measures range and relative velocity directly by Doppler shift; works in rain, fog, dust and darkness. Poor angular resolution and object classification; metal road furniture creates clutter. The backbone of ACC and AEB.
LiDAR: scans a pulsed laser and times the return to build a dense 3-D point cloud; excellent geometry and ranging accuracy (centimetres). Degrades in fog, dust and heavy rain; mechanical spinning units are costly, solid-state units are the route to lower cost, and FMCW units add direct Doppler velocity measurement.
Ultrasonic: cheap, 0.2-5 m, used for parking assist and low-speed obstacle detection; useless at speed.
GNSS (GPS/NavIC) + IMU: absolute position plus dead reckoning through tunnels and urban canyons; RTK or differential correction gets to centimetre level. India's NavIC constellation is being added to automotive receivers.
Proprioceptive sensors: wheel speed, steering angle, yaw rate, longitudinal/lateral accelerometers - shared with ABS and ESC.
Driver monitoring system (DMS): infrared camera watching eye gaze and head pose; mandatory for hands-free Level 2 and Level 3 in several markets.
Redundancy principle: each sensor's weakness is another's strength, so safe systems use diverse sensing (camera + radar minimum) and never a single modality for a safety-critical decision.

The perception - planning - control pipeline

Sensing and pre-processing: raw frames, point clouds and radar detections, time-stamped and corrected for distortion and ego-motion.
Calibration: intrinsic (lens) and extrinsic (sensor-to-vehicle transform) so that all data can be expressed in one vehicle coordinate frame.
Perception: detection, classification and segmentation (convolutional networks, transformers), lane and free-space estimation, traffic sign and signal recognition.
Sensor fusion: combine sources into one consistent world model. Low-level (early) fusion merges raw data; high-level (late) fusion merges each sensor's object lists. Tracking uses the Kalman filter (linear), extended/unscented Kalman filter (non-linear) or a particle filter; association by nearest neighbour or Hungarian algorithm.
Localisation and mapping: match perception against an HD map (lane-level, 10-20 cm accuracy) or build one on the fly - SLAM (simultaneous localisation and mapping).
Prediction: estimate where every other road user will be in the next 1-5 s - the hardest part in Indian mixed traffic with two-wheelers, autos, pedestrians and cattle sharing the lane.
Planning: route (A to B, A*/Dijkstra), behaviour (follow, change lane, yield, stop - finite state machine or learned policy), motion/trajectory (RRT*, lattice planner, polynomial or clothoid path, optimisation with comfort and jerk limits).
Control: track the trajectory. Longitudinal - PID or MPC on throttle and brake; lateral - pure pursuit, Stanley controller or MPC on steering angle. Actuation by drive-by-wire: EPS motor for steering, electronic brake booster/ESC hydraulic unit for braking, electronic throttle or inverter torque command for propulsion.
Actuation and arbitration: a supervisor arbitrates between features (e.g. AEB overrides ACC) and limits authority (torque and jerk limits) so a fault cannot command a violent manoeuvre.

Common ADAS features and how each works

ACC (Adaptive Cruise Control): radar (or camera) measures the gap and closing speed to the lead vehicle and holds a time headway (typically 1-2.2 s selectable) by modulating throttle and brake. Stop-and-go ACC can brake to standstill and resume.
FCW / AEB: computes time to collision; warns first, then pre-charges brakes, then brakes autonomously. City AEB works at low speed, inter-urban AEB at highway speed, VRU AEB detects pedestrians and cyclists.
LDW / LKA / LCA: camera tracks lane markings. LDW warns on crossing; LKA gives a corrective nudge at the boundary; LCA (lane centering) steers continuously to hold the centre.
BSM / RCTA: rear corner radars detect vehicles in the blind spot or crossing behind while reversing.
TSR / ISA: traffic sign recognition by camera (plus map data); intelligent speed assist warns or limits speed to the posted limit.
DAW (driver attention/drowsiness warning): steering-input pattern or DMS camera.
Parking: ultrasonic + surround-view cameras; semi-automatic parking steers while the driver controls speed, fully automatic does both; 360° / bird's-eye view stitches four fisheye cameras.
Foundation electronic aids (Level 0): ABS/CBS, reverse parking sensors, speed warning at 80/120 km/h and seat-belt reminder are legally mandated in India; ESC/ESP is mandated only for certain categories, while traction control and TPMS are fitted commercially but are not mandatory here.

Electrical/electronic architecture, networks and functional safety

Architecture trend: many distributed ECUs → domain controllers (ADAS domain, powertrain, body) → zonal + central compute. A high-performance SoC (NVIDIA Orin, Qualcomm Ride, TI TDA4, Mobileye EyeQ) runs perception; a lock-step safety MCU (Infineon AURIX) supervises.
Networks: CAN / CAN-FD for control messages, LIN for cheap body functions, FlexRay (legacy, time-triggered), Automotive Ethernet (100BASE-T1/1000BASE-T1) for camera and LiDAR bandwidth, SOME/IP and DDS as service protocols, AUTOSAR Classic for real-time ECUs and AUTOSAR Adaptive for high-compute nodes.
Functional safety - ISO 26262 (ISO 26262 series, adopted in India as IS/ISO 26262): hazard analysis and risk assessment rates each hazard by Severity (S0-S3) × Exposure (E0-E4) × Controllability (C0-C3) to give an ASIL from A (lowest) to D (highest); steering and braking functions are typically ASIL C/D. Safety goals flow into redundancy, diagnostics and the safe state.
SOTIF - ISO 21448 (Safety Of The Intended Functionality): covers hazards from performance limits and foreseeable misuse even when nothing is broken - a camera blinded by low sun, a learned model that fails on an unseen object. The dominant concern in ADAS because most ADAS accidents involve no hardware fault.
Cybersecurity: ISO/SAE 21434 for process and UN R155 for CSMS, UN R156 for software updates - relevant to OTA updates.
Verification: MIL → SIL → HIL → vehicle-in-the-loop → proving ground → public road; plus massive scenario-based simulation (CARLA, dSPACE, IPG CarMaker) and re-simulation of recorded drives.

V2X, connectivity and the Indian regulatory picture

V2X: V2V (vehicle), V2I (infrastructure), V2P (pedestrian), V2N (network). Radio options are DSRC/ITS-G5 (IEEE 802.11p) and C-V2X (3GPP LTE-V2X / 5G NR-V2X); India has leaned towards C-V2X. V2X extends awareness beyond line of sight - blind intersection, emergency vehicle warning, platooning.
Indian regulation: road vehicles are regulated under the Motor Vehicles Act, 1988 (as amended by the Motor Vehicles (Amendment) Act, 2019) and the Central Motor Vehicles Rules, 1989; type approval follows AIS (Automotive Industry Standards) and CMVR, tested by ARAI, ICAT, GARC and CIRT. Mandatory items include ABS/CBS, seat-belt reminder, over-speed warning, reverse parking alert, ESC for certain categories, speed limiting devices for commercial vehicles, and AIS 140 (VLT with panic button for public service vehicles); TPMS is covered by AIS 142 but is not mandatory.
Bharat NCAP (operational from 1 October 2023, based on AIS 197) rates adult and child occupant protection and awards Safety Assist points for ESC, seat-belt reminders, pedestrian protection and ADAS features such as AEB and BSD - this is the main commercial driver of ADAS fitment in India.
Legal status of automation in India: there is no Indian regulation permitting Level 3 or above on public roads, and no statutory framework assigning liability to a system rather than a driver; the Motor Vehicles Act still presumes a human driver in control. ADAS features are sold strictly as Level 2 driver aids. Testing is done on private proving grounds and in simulation. Specific AIS numbers for ADAS features (for example AEB and lane-keeping requirements) are being drafted and amended, so quote them only after checking the current ARAI/MoRTH list.
Indian-specific challenges: unmarked or faded lane lines, mixed traffic with two- and three-wheelers, unsignalled intersections, stray animals, unregulated speed breakers, informal road geometry and monsoon visibility - all of which make models trained on Western data perform poorly, so Indian datasets (IDD - India Driving Dataset) are used for retraining.

Advantages and limitations

Advantages: fewer crashes from the human factors that cause most accidents (distraction, fatigue, misjudged gaps); smoother driving, so better fuel economy and range; reduced driver workload; mobility for the elderly and disabled; platooning cuts aerodynamic drag; data for fleet and insurance telematics.
Limitations: cost (sensor set plus compute adds meaningfully to vehicle price); performance limits in bad weather and poor infrastructure; mode confusion and over-trust in Level 2; the long-tail problem - rare scenarios dominate residual risk; validating safety statistically needs billions of kilometres; cybersecurity exposure; unresolved liability and ethics; employment impact on commercial drivers.

Key points

ADAS - electronic systems that assist a driver who remains responsible; AV - a vehicle that performs the whole driving task itself.
SAE J3016 - six levels, 0 to 5; the break between 2 and 3 is who monitors the environment, not how good the steering is.
ODD - the declared conditions in which a feature is valid; outside it the feature must refuse to engage or hand back.
DDT - lateral control + longitudinal control + OEDR; DDT fallback is the response to failure or ODD exit.
Radar - direct range and relative velocity, weather-robust, poor classification; camera - rich semantics, weak in poor light; LiDAR - precise 3-D geometry, costly and weather-sensitive.
Sensor fusion - combining complementary sensors into one world model; early fusion merges raw data, late fusion merges object lists.
Kalman filter - recursive predict-and-correct estimator for tracking; EKF/UKF handle non-linear motion and measurement models.
Pipeline - sense → perceive → fuse → localise → predict → plan → control → actuate, repeated typically at 10-100 Hz.
Lateral controllers - pure pursuit (geometric look-ahead), Stanley (cross-track + heading error), MPC (constrained optimisation over a horizon).
TTC (time to collision) - the core trigger for FCW and AEB; warning thresholds are around 2-2.5 s, autonomous braking well under 1.5 s.
Time headway - the gap setting in ACC, in seconds, so the distance scales with speed.
Drive-by-wire - electronic actuation of steering (EPS), braking (electronic booster) and propulsion; prerequisite for any automation, and native to EVs.
ISO 26262 / ASIL A-D - functional safety for faults; ASIL from Severity × Exposure × Controllability.
ISO 21448 (SOTIF) - safety when nothing is broken but the function's performance is insufficient; central to ADAS.
ISO/SAE 21434, UN R155/R156 - cybersecurity management and software update management.
V2X - V2V/V2I/V2P/V2N; DSRC (802.11p) vs C-V2X (LTE/5G), India leaning to C-V2X.
Indian law - Motor Vehicles Act, 1988 and CMVR 1989 presume a human driver; no framework yet for Level 3+ on public roads; Bharat NCAP (from 1 Oct 2023) rewards ADAS in Safety Assist scoring.
EVs and automation converge - by-wire actuation, large 12/48 V electrical capacity and software-defined architecture make EV platforms the natural host for ADAS.

Formulas and examples

TTC = relative distance / relative closing speed = Δd / Δv (s; Δd in m, Δv in m/s, constant-velocity assumption)
Safe following distance d = v × τ + d₀ (τ = time headway in s, v = ego speed in m/s, d₀ = standstill gap in m)
Stopping distance s = v·t_r + v²/(2·a) (t_r = reaction/system latency in s, a = deceleration in m/s², v in m/s)
Radar Doppler shift f_d = 2 v_r f₀ / c (v_r = relative radial velocity m/s, f₀ = carrier Hz, c = 3×10⁸ m/s)
LiDAR/ultrasonic range R = c·t/2 (light) or R = v_sound·t/2 (sound); v_sound ≈ 343 m/s at 20 °C
Ackermann steering: δ ≈ L/R (δ = steer angle rad, L = wheelbase m, R = turn radius m)
Pure pursuit curvature κ = 2·sin(α)/l_d → δ = atan(2 L sin α / l_d) (α = angle to look-ahead point, l_d = look-ahead distance m)
Stanley steering δ = θ_e + atan(k·e / v) (θ_e = heading error rad, e = cross-track error m, k = gain, v = speed m/s)
Lateral acceleration a_y = v²/R (comfort limit about 1.5-3 m/s²)
Kalman update: x̂ = x̄ + K(z − H x̄); K = P̄Hᵀ(H P̄ Hᵀ + R)⁻¹ (P = covariance, R = measurement noise, z = measurement)
Camera depth (stereo) Z = f·B/d (f = focal length in pixels, B = baseline m, d = disparity in pixels)

Example 1 - TTC and AEB decision. Ego at 80 km/h closes on a truck at 50 km/h, gap 40 m.

v_ego = 80/3.6 = 22.22 m/s, v_truck = 50/3.6 = 13.89 m/s → Δv = 8.33 m/s.
TTC = 40 / 8.33 = 4.8 s → beyond the warning threshold, so ACC simply decelerates gently.
If instead the truck is stationary: Δv = 22.22 m/s, TTC = 40/22.22 = 1.8 s → FCW fires; braking at 8 m/s² after 0.3 s latency needs 22.22×0.3 + 22.22²/(2×8) = 6.67 + 30.86 = 37.5 m, just inside 40 m, so a full-authority AEB stop still avoids impact with about 2.5 m to spare.

Example 2 - ACC headway. Time headway set to 1.8 s, standstill gap 4 m.

At 60 km/h (16.67 m/s): d = 16.67×1.8 + 4 = 34.0 m.
At 100 km/h (27.78 m/s): d = 27.78×1.8 + 4 = 54.0 m.
At 1.0 s headway and 100 km/h: d = 27.78 + 4 = 31.8 m - shows why a short headway is unsafe: at a 0.5 s system reaction plus 7 m/s² braking the ego needs 27.78×0.5 + 27.78²/14 = 13.9 + 55.1 = 69.0 m to stop from 100 km/h.

Example 3 - Radar Doppler. A 77 GHz radar sees a target closing at 30 m/s.

f_d = 2 × 30 × 77×10⁹ / 3×10⁸ = 2 × 30 × 256.7 = 15,400 Hz ≈ 15.4 kHz.
Range from round-trip time 0.667 µs: R = 3×10⁸ × 0.667×10⁻⁶ / 2 = 100 m.

Example 4 - Lateral control and comfort. A lane change of 3.5 m lateral offset completed in 4 s at 90 km/h (25 m/s).

Using a sinusoidal lateral acceleration profile, peak a_y ≈ 2π·Y/T² = 6.28 × 3.5/16 = 1.4 m/s² - inside the comfort limit; a hurried 2 s lane change would need 6.28 × 3.5/4 = 5.5 m/s², far above it, so the planner holds the manoeuvre to 4 s or more.
Equivalent steady-radius check: holding a_y = 1.4 m/s² at 25 m/s means R = v²/a_y = 625/1.4 ≈ 446 m; with a 2.7 m wheelbase, δ ≈ 2.7/446 = 0.0061 rad = 0.35° at the road wheels - roughly 5-6° at the steering wheel for a 16:1 ratio.

Example 5 - Stereo depth resolution. Stereo camera, f = 1,200 px, baseline B = 0.25 m.

Disparity for a car at 30 m: d = 1,200 × 0.25/30 = 10 px.
At 60 m: d = 5 px. A ±0.5 px disparity error at 60 m gives Z between 1,200×0.25/5.5 = 54.5 m and 1,200×0.25/4.5 = 66.7 m - a ±6 m uncertainty, which is why radar (direct ranging) is fused in for longitudinal control.

Sensor	Measures	Typical range	Weather	Cost	Main use
Mono camera	Colour, texture, classification	50-150 m	Poor in glare, fog, night	Low	Lanes, signs, VRU
Stereo camera	Above + depth	50-80 m	Poor	Medium	Free space, depth
Radar (77 GHz)	Range, relative velocity	150-250 m	Excellent	Medium	ACC, AEB, BSM
LiDAR	3-D geometry	100-250 m	Poor in fog/dust	High	Mapping, L4 perception
Ultrasonic	Short range	0.2-5 m	Good	Very low	Parking
GNSS + IMU	Position, attitude	Global	Blocked in tunnels	Low-High	Localisation

Aspect	Level 2	Level 3
Who monitors the road	Human driver	The system
Hands/eyes	Eyes on road required	Eyes may be off within ODD
Fallback	Driver, immediately	Driver, after a request to intervene
Legal driver	Human	System within ODD (where law permits)
Indian road legality	Sold today	Not permitted on public roads

Where it is used

OEM and Tier-1 engineering: Bosch, Continental, ZF, Aptiv, Valeo, Denso and Mobileye India centres develop radar, camera and domain-controller software; Mahindra, Tata Motors, Maruti Suzuki and Hyundai India integrate and calibrate Level 2 suites.
Automotive software services: KPIT Technologies, Tata Elxsi, L&T Technology Services, Cyient, HCLTech and Capgemini Engineering supply AUTOSAR, perception, HIL and ISO 26262 work to global programmes - the largest ADAS recruiter group in India.
Testing and homologation: ARAI Pune, ICAT Manesar, GARC Oragadam and NATRAX Indore run proving-ground, HIL and type-approval testing for CMVR/AIS and Bharat NCAP.
Commercial fleets and mining: collision warning and driver-fatigue systems on trucks and buses, AIS 140 telematics on public service vehicles, and autonomous haul trucks in mines and ports.
Agriculture and off-road: auto-steer and GNSS guidance on Mahindra and Escorts/John Deere tractors - a genuine Level 2-4 application already deployed in Indian fields.
Job roles: ADAS validation engineer, perception/computer vision engineer, sensor fusion engineer, functional safety (ISO 26262) engineer, HIL/test-bench engineer, AUTOSAR developer, calibration engineer, data annotation lead.

Interview questions

Q: What is the real difference between Level 2 and Level 3? A: At Level 2 the human monitors the driving environment and is always the driver; at Level 3 the system monitors it and drives within its ODD, and the human only takes over after a request to intervene.
Q: Why is radar used for ACC rather than a camera alone? A: Radar measures range and relative velocity directly via Doppler and works in rain, fog and darkness, whereas camera range is inferred and degrades badly in poor visibility.
Q: What is an ODD and why does it matter legally? A: It is the declared set of conditions in which a feature is valid; safety claims, type approval and liability all apply only inside it, and the system must hand back or refuse to engage outside it.
Q: Difference between ISO 26262 and ISO 21448? A: ISO 26262 handles hazards from malfunctions and random/systematic faults (ASIL); ISO 21448 (SOTIF) handles hazards when the system is fault-free but its performance or the user's misuse is insufficient.
Q: What is sensor fusion and why not just use the best single sensor? A: Fusion merges complementary sensors into one world model; no single modality covers semantics, direct velocity and precise 3-D geometry, and redundancy is required for a safety-critical decision.
Q: How does AEB decide when to brake? A: It estimates time to collision from relative distance and closing speed, warns at roughly 2-2.5 s, pre-fills the brakes, and applies full braking inside about 1 s if the driver does not respond.
Q: What does a Kalman filter do in a tracker? A: It predicts each object's state from a motion model and corrects it with the new measurement, weighting the two by their covariances, giving a smoothed estimate and a velocity even from noisy position data.
Q: Compare pure pursuit and the Stanley controller. A: Pure pursuit steers geometrically towards a look-ahead point on the path using rear-axle geometry; Stanley uses front-axle cross-track error plus heading error and tracks more accurately at low speed but is more sensitive to noise.
Q: Why are ADAS and EVs discussed together? A: EVs are already by-wire and software-defined with ample electrical power and a central compute architecture, so they host automation with far less retrofitting than a mechanically controlled IC vehicle.
Q: Can a Level 3 car be legally driven on Indian roads today? A: No - the Motor Vehicles Act, 1988 and CMVR presume a human driver in control and there is no Indian framework approving Level 3 or assigning liability to a system; features sold in India are Level 2 driver aids.
Q: What makes Indian roads harder for ADAS than European roads? A: Faded or absent lane markings, dense mixed traffic of two- and three-wheelers and pedestrians, unsignalled junctions, unmarked speed breakers, animals and monsoon visibility - so models trained on Western data must be retrained on Indian datasets.
Q: What is "mode confusion" and why is it a Level 2 safety problem? A: The driver misjudges what the system is doing or how much it can handle, over-trusts it and stops supervising - the cause of several Level 2 crashes, which is why driver monitoring and clear HMI are required.
Q: Why is Automotive Ethernet replacing CAN in ADAS? A: Cameras and LiDAR need tens to hundreds of Mbit/s, far beyond CAN-FD's few Mbit/s; CAN still carries low-bandwidth control and diagnostic messages.

Common mistakes

Calling any ADAS car "self-driving" or "autonomous" - Level 2 features are assistance and the driver is legally and practically in charge.
Describing the levels by hardware or capability instead of by who monitors the environment and who is the fallback.
Treating ADAS as "AI only" and ignoring calibration, latency, ASIL decomposition, fail-operational design and validation, which is where most engineering effort actually goes.
Mixing units in TTC and braking sums - using km/h with metres and seconds, or forgetting to add reaction/system latency to the braking distance.
Assuming more sensors automatically means safer: without time synchronisation, calibration and a fusion strategy, extra sensors add false positives.
Confusing ISO 26262 (faults) with ISO 21448 (performance limitations), or quoting an ASIL without the S × E × C reasoning.
Quoting a withdrawn or imagined AIS/IS number for an ADAS feature - the ADAS-specific Indian standards are still being drafted and amended, so name the requirement rather than guess a number.