Important bug - static box when we lose track

[PetropoulakisPanagiotis]

I found and validated a bug from my side.

Have a look at this part of the code:

        for trk in reversed(self.trackers):
            if trk.last_observation.sum() > 0:
                d = trk.last_observation[:4]
            else:
                d = trk.get_state()[0]

The self.last_observation is set to "-1" values only during the def __init__:
self.last_observation = np.array([-1, -1, -1, -1, -1]) # placeholder
This means in the code I have pasted above, the if statement
if trk.last_observation.sum() > 0:, will never be true.

In my local side, I observed that once we loose track of an object, we use always d = trk.last_observation[:4], and this means that the tracked box is always the same. I could see this in a video I generated. When we lose the track, I can see that the box remains static, which is not ideal.

I did this hack to make it work, but I am not sure if it is optimal, you can also suggest a fix please:

    def update(self, bbox):
        """
        Updates the state vector with observed bbox.
        """
        if bbox is not None:
            if self.last_observation.sum() >= 0:  # no previous observation
                previous_box = None
                for i in range(self.delta_t):
                    dt = self.delta_t - i
                    if self.age - dt in self.observations:
                        previous_box = self.observations[self.age-dt]
                        break
                if previous_box is None:
                    previous_box = self.last_observation
                """
                  Estimate the track speed direction with observations \Delta t steps away
                """
                self.velocity = speed_direction(previous_box, bbox)
            
            """
              Insert new observations. This is a ugly way to maintain both self.observations
              and self.history_observations. Bear it for the moment.
            """
            self.last_observation = bbox
            self.observations[self.age] = bbox
            self.history_observations.append(bbox)

            self.time_since_update = 0
            self.history = []
            self.hits += 1
            self.hit_streak += 1
            self.kf.update(convert_bbox_to_z(bbox))
        else:
            self.last_observation = np.array([-1, -1, -1, -1, -1])
            self.kf.update(bbox)