Reliable estimation of a vehicle's ground-relative speed is a prerequisite for many automated functions on construction equipment, yet the dominant signals available today fail in the conditions where construction machines operate, wheel RPM degrades on slip-prone surfaces and GPS degrades under obstructed sky. This thesis investigates whether a single, fixed-pitch monocular camera can fill that gap and under which conditions it can do so with useful accuracy.
Three optical-flow methods that span the main families used in practice are compared inside a common speed-estimation pipeline: a sparse classical method (Lucas-Kanade), a dense classical method (Farneback) and a dense learning-based method (RAFT). Each method feeds the same downstream stages, which unproject image-plane motion onto an assumed ground plane and filters out unreliable flow vectors. The pipeline is evaluated on the public Comma2k19 driving dataset, with the vehicle's own CAN-bus speed signal serving as the reference.
By holding the pipeline fixed and varying only the optical-flow stage, the experiments isolate the contribution of the flow algorithm itself across daytime highway driving, night driving and segments with uneven terrain. Lucas-Kanade produced the most accurate and most stable speed estimates overall. Farneback was the most resilient at night, but at an accuracy too low to be useful in practice. RAFT was competitive only on a single carefully chosen ideal segment and degraded sharply at night and on uneven terrain. None of the methods produced reliably usable speed under low-light or strongly disturbed conditions, which identifies the principal failure modes. The evaluation uses passenger-car highway data (Comma2k19); the results therefore establish feasibility under daylight highway conditions and are not direct evidence of performance on construction machines.