class LaneHypothesis(object): def __init__(self, position, momentum, weight=1.0): self.positions = [] self.positions.append(position) self.weight = weight self.prev_data = None self.momentum = momentum self.dirfilter = None def _feat_probability(self, feat): std = 0.1 return scipy.stats.norm.pdf(1.0 - feat, 0.0, scale=std) @property def position(self): return self.positions[-1] def evidence(self, data): if self.dirfilter is None: self.dirfilter = AngleFilter(data.shape) c = self.dirfilter.get_component_values(data, mode='valid')[0][0] feat = self.dirfilter.basis.max_response_value(c)/115.0 #feat = pearsonr(data.flatten(), self.prev_data.flatten())[0] #prob = self._feat_probability(lightness*sobeldiff) return feat def prior(self, prev): disp = self.position - prev.position self.momentum = 0.5*disp + 0.5*prev.momentum prob = scipy.stats.norm.pdf(disp, loc=self.momentum, scale=20.0) return prob
def evidence(self, data): if self.dirfilter is None: self.dirfilter = AngleFilter(data.shape) c = self.dirfilter.get_component_values(data, mode='valid')[0][0] feat = self.dirfilter.basis.max_response_value(c)/115.0 #feat = pearsonr(data.flatten(), self.prev_data.flatten())[0] #prob = self._feat_probability(lightness*sobeldiff) return feat
def detwalk(image, prev_start): blocksize = (9, 9) filt = AngleFilter(blocksize) rows = range(0, image.shape[0]-blocksize[1], blocksize[1]/2) tracker = HypothesisTracker() particles = [] for i, row in enumerate(rows): candidates = [] d = image[row:row+blocksize[1]] comp = filt.get_component_values(d, mode='valid')[0] angles = filt.basis.max_response_angle(comp.T) values = filt.basis.max_response_value(comp.T) #valid = values > 1.5 valid = np.argsort(values)[::-1][:5] indices = np.array(range(len(angles)), dtype=float) angles = np.array(angles[valid]) feats = np.array((indices[valid])) weights = np.array((values[valid])) if len(feats) < 2: continue feats = feats.reshape((len(feats), 1)) clusters = scipy.cluster.hierarchy.fclusterdata(feats, 3, criterion='distance') for c in np.unique(clusters): x = np.mean(feats[clusters == c])+blocksize[0]/2.0 y = row+blocksize[1]/2.0 #pylab.scatter(x, y) candidates.append(((x, y), np.median(angles[clusters == c]), np.mean(values[clusters == c]))) candidates.sort(key=lambda x: x[-1]) particles = tracker.spawn_hypotheses(particles, candidates) particles.sort(key=lambda p: p.mean_weight(), reverse=True) return particles
