embedded - how to use accelerometer sensor to detect and match two toy cars' collision -

i using accelerometer sensor collect toy car's acceleration data tell whether bumps or collide toy car. idea checking whether there sharp change of acceleration data (for simplicity, considering x , y axis acceleration data since toy car running on smooth ground). when car started stop, there dramatic change acceleration data. need distinguish them. based on instinct, bumping cause sharp changed acceleration in short time, while engine starting cause smooth changed acceleration data. think should quite common question. new hand. i'd know whether there algorithm or method distinguish them. great if there implementation in c. lot.

[update bumping situations]
let me make question more specific. project involves several toy cars, can bump each other, or bump other things(such walls). need find out 2 cars collided each other. idea try compare 2 cars' acceleration data see if matched same pattern. divided project several steps:
1. detect bumps, need distinguish car bumping being started or other situations, such jolts on road.
2. match 2 cars' bump features see if bumped each other. kind of features can use?

[update]

found url containing useful information peak finding.
http://terpconnect.umd.edu/~toh/spectrum/peakfindingandmeasurement.htm

i post answer because there @ least interesting elements question close appropriate stackoverflow.

1. detect bumps, need distinguish car bumping being started or other situations, such jolts on road.

a collision involves large negative acceleration of high magnitude. normal starting positive, while stopping of low magnitude.

moreover use motor control in fusion accelerometer. if acceleration matches being demanded motors, normal, while if acceleration uncorrelated motor demand, independent , caused external force.

acceleration caused uneven terrain (i deliberately avoiding terms "bumps" , "jolts" confusing , imprecise) has vertical component , short-term such in windowed integrator integrates 0 (i.e. terrain "flat" on average). such accelerations damped suspension vehicle may have, handling vehicle specific, ignoring vertical component may necessary.

2. match 2 cars' bump features see if bumped each other. kind of features can use?

if 2 vehicles share common or synchronised time source, such gps or common remote controller, both vehicles experiencing abnormal acceleration @ precisely same time indication collided each other. head-on collision of vehicles same mass have broadly equal acceleration profiles. oblique collisions more complex (and more likely), involve larger acceleration component perpendicular intended direction of travel (i.e. sideways). might correlate sideways acceleration steering demand in similar manner motor demand. if there sideways movement without steering demand, external force.

i think overall question requires empirical data. need run experiments produce kind of situations wish distinguish, capture data, plot , observe distinguishing characteristics each situation. data should include accelerometer axes, plus motor demand. can determine how may distinguished mathematically. can build prospective algorithms , pass captured data through them verify work before implementing them on vehicles.

i believe in case solution involve elements of digital signal processing - unlikely decision can made in discrete instantaneous inputs.

since nature collision largely non-deterministic , variable, simpler model "normal" driving situations on range of terrain, , detect outside normal potential collision. rather try characterise collision, characterise "normal" , detect "abnormal".