Trajectory Search

Search mode: retrieval by how the ego vehicle moved, either the shape of its path or statistics over that motion.

Trajectory search retrieves clips using ego-vehicle motion. It supports two modes: shape search, which finds clips whose ego path is geometrically similar to a reference clip, and pattern search, which filters clips by evaluating expressions over motion statistics such as speed, acceleration, and curvature. Shape search is a ranked mode; pattern search is filter-only. Both require ego trajectory data.

Reach for it when the manoeuvre matters more than the scenery. Use shape search when you have a reference clip with an interesting manoeuvre and want geometrically similar paths. Matching is speed-invariant, so the same curve driven slowly still qualifies. Use pattern search to filter by a named behaviour such as hard braking, swerving, or a prolonged stop.

Reach for something else when what you care about is in the scene rather than the ego motion, or when the dataset carries no ego trajectories, since neither mode returns anything without them.

Trajectory shape search: clips with similar geometric path to a reference clip.

Shape Search

Finds clips whose ego path has a similar geometric shape to a reference clip. Trajectories are normalised to their starting position, so the comparison is purely about shape: the same curve driven at different speeds will match. A time window can be specified to restrict the match to a sub-segment, allowing search for a specific manoeuvre without requiring the full path to match. Shape search is a ranked mode.

Pattern Search

Pattern search filters clips by evaluating expressions over per-clip motion statistics. For each clip, the following scalar arrays are precomputed from the ego trajectory, with one value per timestep sampled at 10 Hz:

Variable	Description
speed	Ego speed in m/s
speed_kph	Ego speed in km/h
acceleration	Longitudinal acceleration in m/s² (negative = braking)
jerk	Rate of change of acceleration in m/s³
curvature	Path curvature at each timestep (higher = sharper turn)

Expressions can reference any of these arrays using standard mathematical and logical operations. This is equivalent to writing a Python expression in an interactive notebook: the expression is evaluated against each clip’s arrays and the clip is included in results if the expression evaluates to True. Aggregate functions such as sum(), mean(), max(), min(), and any() are available, as are element-wise comparisons.

Examples:

mean(speed) > 20
max(curvature) > 0.2 and mean(speed) > 10
sum(acceleration < -3.0) > 10
any(speed < 0.5) and any(speed_kph > 50)

Predefined Patterns

The following named patterns are available as shortcuts in the UI. Selecting one is equivalent to entering the corresponding expression:

Pattern name	Expression	Description
high_curvature	sum(curvature > 0.15) > 10	Sharp turning sustained over at least 1 second
stop_go	any(speed < 0.5) and any(speed > 3.0) and min(where(speed > 3.0)[0]) > min(where(speed < 0.5)[0])	Speed transitions from near-stop to moving (stop precedes go)
hard_braking	sum(acceleration < -3.0) > 10	Strong deceleration sustained over at least 1 second
prolonged_stop	sum(speed < 0.5) > 150	Stationary for more than 15 seconds
idle_to_cruise	any(speed < 0.5) and any(speed > 10.0) and min(where(speed > 10.0)[0]) > min(where(speed < 0.5)[0])	Clip starts stationary and reaches open-road speed (idle precedes cruise)
high_speed_swerve	sum(curvature > 0.2) > 10 and sum(speed_kph > 50) > 10	Sharp lateral movement at highway speed
moving_ego	sum(speed_kph > 5) > 10	Ego vehicle is in motion for at least 1 second

Custom expressions can be entered directly in the Custom motion filter field and support the same syntax and variables as the predefined patterns.