Merge 719b9e8 into 07297fe

carball/analysis/events/bump_detection/bump_analysis.py

@@ -1,37 +1,292 @@
+import itertools
+import logging
+from typing import Dict
+
+import numpy as np
 import pandas as pd
+from carball.generated.api.player_pb2 import Player
+
 from carball.generated.api.stats.events_pb2 import Bump
 
 from carball.generated.api.player_id_pb2 import PlayerId
 from carball.json_parser.game import Game
 
 from carball.generated.api import game_pb2
 
+logger = logging.getLogger(__name__)
+
+# Decreasing this, risks not counting bumps where one car is directly behind another (driving in the same direction).
+# Increasing this, risks counting non-contact close proximity (e.g. one car cleanly jumped over another =/= bump).
+PLAYER_CONTACT_DISTANCE = 200
+
+# Needs to be relatively high to account for two cars colliding 'diagonally': /\
+MAX_BUMP_ALIGN_ANGLE = 60
+
+# Currently arbitrary:
+MIN_BUMP_VELOCITY = 5000
+
+# Approx. half of goal height.
+# (could be used to discard all aerial contact as bumps, although rarely an aerial bump WAS, indeed, intended)
+AERIAL_BUMP_HEIGHT = 300
 
+
+# TODO Post-bump analysis // Bump impact analysis.
 class BumpAnalysis:
     def __init__(self, game: Game, proto_game: game_pb2):
         self.proto_game = proto_game
 
-    def get_bumps_from_game(self, data_frame: pd.DataFrame):
+    def get_bumps_from_game(self, data_frame: pd.DataFrame, player_map):
         self.create_bumps_from_demos(self.proto_game)
-
-        self.analyze_bumps(data_frame)
+        self.create_bumps_from_player_proximity(data_frame, player_map)
 
     def create_bumps_from_demos(self, proto_game):
         for demo in proto_game.game_metadata.demos:
             self.add_bump(demo.frame_number, demo.victim_id, demo.attacker_id, True)
 
+    def create_bumps_from_player_proximity(self, data_frame: pd.DataFrame, player_map: Dict[str, Player]):
+        """
+        Attempt to find all instances between each possible player combination
+        where they got within PLAYER_CONTACT_DISTANCE.
+        Then, add each instance to the API.
+
+        NOTES:
+        Currently, this yields more 'bumps' than there are actually.
+          This is mostly due to aerial proximity, where a bump was NOT intended or no contact was made.
+          This also occurs near the ground, where cars flip awkwardly past each other.
+        """
+
+        # An array of player names to get player combinations; and a dict of player names to their IDs to create bumps.
+        player_names = []
+        player_name_to_id = {}
+        for player in player_map.values():
+            player_names.append(player.name)
+            player_name_to_id[player.name] = player.id
+
+        # For each player pair combination (nCr), get all frames where they got close and then filter those as bumps.
+        for player_pair in itertools.combinations(player_names, 2):
+            players_close_frame_idxs = BumpAnalysis.get_players_close_frame_idxs(data_frame,
+                                                                                 str(player_pair[0]),
+                                                                                 str(player_pair[1]))
+
+            if len(players_close_frame_idxs) > 0:
+                likely_bumps = BumpAnalysis.filter_bumps(data_frame, player_pair, players_close_frame_idxs)
+                self.add_non_demo_bumps(likely_bumps, player_name_to_id)
+            else:
+                logger.info("Players (" + player_pair[0] + " and " + player_pair[1] + ") did not get close "
+                                                                                      "during the match.")
+
     def add_bump(self, frame: int, victim_id: PlayerId, attacker_id: PlayerId, is_demo: bool) -> Bump:
+        """
+        Add a new bump to the proto_game object.
+        """
         bump = self.proto_game.game_stats.bumps.add()
         bump.frame_number = frame
         bump.attacker_id.id = attacker_id.id
         bump.victim_id.id = victim_id.id
         if is_demo:
             bump.is_demo = True
 
-    def analyze_bumps(self, data_frame:pd.DataFrame):
-        for bump in self.proto_game.game_stats.bumps:
-            self.analyze_bump(bump, data_frame)
+    def add_non_demo_bumps(self, likely_bumps, player_name_to_id):
+        """
+        Add a new bump to the proto_game object.
+        This method takes an array of likely (filtered) bumps, in the following form:
+            (frame_idx, attacker_name, victim_name)
+        and carefully adds them to the proto_game object (i.e. check for demo duplicates).
+        """
+
+        # Get an array of demo frame idxs to compare later.
+        demo_frame_idxs = []
+        for demo in self.proto_game.game_metadata.demos:
+            demo_frame_idxs.append(demo.frame_number)
+
+        # For each bump tuple, if its frame index is not similar to a demo frame index, add it via add_bump().
+        for likely_bump in likely_bumps:
+            likely_bump_frame_idx = likely_bump[0]
+            if not any(np.isclose(demo_frame_idxs, likely_bump_frame_idx, atol=10)):
+                self.add_bump(likely_bump[0], player_name_to_id[likely_bump[2]], player_name_to_id[likely_bump[1]],
+                              is_demo=False)
+
+    @staticmethod
+    def filter_bumps(data_frame, player_pair, players_close_frame_idxs):
+        """
+        Filter the frames where two players got close - the filtered frames are likely bumps.
+
+        The main principle used is the angle between two vectors (aka 'alignment'):
+            the velocity vector of player A;
+            the positional vector of the difference between the positions of player B and player A.
+        Both of these vectors point away from player A, and if the angle between them is small - it is likely that
+        Player A bumped Player B. (Velocity going 'through' Player B's Position)
+
+        Some further checks are done to categorise the bump (i.e. is_aerial_bump(), is_bump_velocity() )
+        """
+        likely_bumps = []
+
+        # Split a list of frame indexes into intervals where indexes are within 3 of each other (i.e. consecutive).
+        players_close_frame_idxs_intervals = BumpAnalysis.get_players_close_intervals(players_close_frame_idxs)
+
+        # For each such interval, take (currently only) the first frame index and analyse car behaviour.
+        for interval in players_close_frame_idxs_intervals:
+            frame_before_bump = interval[0]
+
+            # Calculate both player bump alignments (see comment at method top).
+            p1_alignment_before = BumpAnalysis.get_player_bump_alignment(data_frame, frame_before_bump,
+                                                                         player_pair[0], player_pair[1])
+            p2_alignment_before = BumpAnalysis.get_player_bump_alignment(data_frame, frame_before_bump,
+                                                                         player_pair[1], player_pair[0])
+
+            # Determine the attacker and the victim (see method for more info). is_ambiguous signifies whether
+            # the attacker-victim pair is clear or not.
+            attacker, victim, is_ambiguous = BumpAnalysis.determine_attacker_victim(player_pair[0], player_pair[1],
+                                                                               p1_alignment_before, p2_alignment_before)
+
+            # Determine if the bump was above AERIAL_BUMP_HEIGHT.
+            is_aerial_bump = BumpAnalysis.is_aerial_bump(data_frame, player_pair[0], player_pair[1], frame_before_bump)
+
+            # Append the current bump data to likely bumps, if there is an attacker and a victim
+            # and if it wasn't an aerial bump (most often it isn't intended, and there is often awkward behaviour).
+            if attacker is not None and victim is not None and not is_aerial_bump:
+                likely_bump = (frame_before_bump, attacker, victim)
+                likely_bumps.append(likely_bump)
+
+            # NOT YET IMPLEMENTED: Check if interval is quite long - players may be in rule 1 :) or might be a scramble.
+            # BumpAnalysis.analyse_prolonged_proximity(data_frame, interval, player_pair[0], player_pair[1])
+
+        return likely_bumps
+
+    @staticmethod
+    def get_player_bump_alignment(data_frame, frame_idx, p1_name, p2_name):
+        """
+        Calculate and return the angle between:
+            the velocity vector of player A;
+            the positional vector of the difference between the positions of player B and player A.
+        """
+
+        # Get the necessary data from the DataFrame at the given frame index.
+        p1_vel_df = data_frame[p1_name][['vel_x', 'vel_y', 'vel_z']].loc[frame_idx]
+        p1_pos_df = data_frame[p1_name][['pos_x', 'pos_y', 'pos_z']].loc[frame_idx]
+        p2_pos_df = data_frame[p2_name][['pos_x', 'pos_y', 'pos_z']].loc[frame_idx]
+
+        # Get the distance vector, directed from p1 to p2.
+        # Then, convert it to a unit vector.
+        pos1_df = p2_pos_df - p1_pos_df
+        pos1 = [pos1_df.pos_x, pos1_df.pos_y, pos1_df.pos_y]
+        unit_pos1 = pos1 / np.linalg.norm(pos1)
+
+        # Get the velocity vector of p1.
+        # Then, convert it to a unit vector.
+        vel1 = [p1_vel_df.vel_x, p1_vel_df.vel_y, p1_vel_df.vel_z]
+        unit_vel1 = vel1 / np.linalg.norm(vel1)
+
+        # Find the angle between the positional vector and the velocity vector.
+        # NOTE: This is currently converted to DEGREES, not sure if this is bad..? ( - DivvyC)
+        ang = (np.arccos(np.clip(np.dot(unit_vel1, unit_pos1), -1.0, 1.0))) * 180 / np.pi
+        return ang
+
+    @staticmethod
+    def get_players_close_frame_idxs(data_frame, p1_name, p2_name):
+        """
+        For a pair of players, find all frame indexes where they got within PLAYER_CONTACT_DISTANCE of each other.
+        Note that they did NOT necessarily make contact.
+        """
+
+        # Separate the positional data of each given player from the full DataFrame and lose the NaN value rows.
+        p1_pos_df = data_frame[p1_name][['pos_x', 'pos_y', 'pos_z']].dropna(axis=0)
+        p2_pos_df = data_frame[p2_name][['pos_x', 'pos_y', 'pos_z']].dropna(axis=0)
+
+        # Calculate the vector distances between the players, and store them as a pd.Series (1D DataFrame).
+        distances = (p1_pos_df.pos_x - p2_pos_df.pos_x) ** 2 + \
+                    (p1_pos_df.pos_y - p2_pos_df.pos_y) ** 2 + \
+                    (p1_pos_df.pos_z - p2_pos_df.pos_z) ** 2
+        distances = np.sqrt(distances)
+
+        # Only keep values < PLAYER_CONTACT_DISTANCE (see top of class).
+        players_close_series = distances[distances < PLAYER_CONTACT_DISTANCE]
+        # Get the frame indexes of the values (as an ndarray).
+        players_close_frame_idxs = players_close_series.index.to_numpy()
+        return players_close_frame_idxs
+
+    @staticmethod
+    def get_players_close_intervals(players_close_frame_idxs):
+        """
+        Separate a list of frame indexes into intervals with consecutive frame indexes.
+        E.g. [3, 4, 5, 7, 19, 21, 23, 24, 57] is turned into [[3, 4, 5, 7], [21, 23, 24], [57]]
+        """
+
+        all_intervals = []
+        interval = []
+        for index, frame_idx in enumerate(players_close_frame_idxs):
+            diffs = np.diff(players_close_frame_idxs)
+            interval.append(frame_idx)
+            if index >= len(diffs) or diffs[index] >= 3:
+                all_intervals.append(interval)
+                interval = []
+        return all_intervals
+
+    @staticmethod
+    def determine_attacker_victim(p1_name, p2_name, p1_alignment, p2_alignment):
+        """
+        Try to 'guesstimate' the attacker and the victim by comparing bump alignment angles.
+            If both bump alignments are above MAX_BUMP_ALIGN_ANGLE, both values are None (no solid attacker/victim)
+            If both bump alignments are within 45deg of each other, both values are None (both attackers)
+
+        :return: A tuple in the form (Attacker, Victim, T) or (None, None, T/F), where the last bool signifies whether
+        the attacker/victim are ambiguous (T) or not (F).
+        """
+
+        if abs(p1_alignment) < MAX_BUMP_ALIGN_ANGLE or abs(p2_alignment) < MAX_BUMP_ALIGN_ANGLE:
+            # if abs(p1_alignment - p2_alignment) < 45:
+            #     # TODO Rework? This would indicate that the bump is ambiguous (no definite attacker/victim).
+            #     return p1_name, p2_name, True
+            if p1_alignment < p2_alignment:
+                return p1_name, p2_name, False
+            elif p2_alignment < p1_alignment:
+                return p2_name, p1_name, False
+
+        # This is ambiguous - neither player had an attacking bump angle.
+        return None, None, True
+
+    @staticmethod
+    def analyse_prolonged_proximity(data_frame, interval, p1_name, p2_name):
+        # TODO Redo this to do some proper analysis.
+        if len(interval) > 10:
+            print(" > Scramble between " + p1_name + " and " + p2_name)
+        # NOTE: Could try analysing immediate post-bump effects.
+        # elif len(interval) >= 5:
+        #     frame_after_bump = interval[len(interval) - 1]
+        #     p1_alignment_after = BumpAnalysis.get_player_bump_alignment(data_frame, frame_after_bump,
+        #                                                                 p1_name, p2_name)
+        #     p2_alignment_after = BumpAnalysis.get_player_bump_alignment(data_frame, frame_after_bump,
+        #                                                                 p2_name, p1_name)
+
+    @staticmethod
+    def is_aerial_bump(data_frame: pd.DataFrame, p1_name: str, p2_name: str, at_frame: int):
+        p1_pos_z = data_frame[p1_name].pos_z.loc[at_frame]
+        p2_pos_z = data_frame[p2_name].pos_z.loc[at_frame]
+        if all(x > AERIAL_BUMP_HEIGHT for x in [p1_pos_z, p2_pos_z]):
+            # if all(abs(y) > 5080 for y in [p1_pos_y, p2_pos_y]):
+            #     print("Backboard bump?")
+            return True
+        else:
+            return False
 
+    @staticmethod
+    def is_bump_alignment(bump_angles):
+        # Check if all bump alignment angles in the first half of the interval are above MAX_BUMP_ALIGN_ANGLE.
+        if all(x > MAX_BUMP_ALIGN_ANGLE for x in bump_angles):
+            return False
+        else:
+            return True
 
-    def analyze_bump(self, bump: Bump, data_frame:pd.DataFrame):
-        frame_number = bump.frame_number
+    @staticmethod
+    def is_bump_velocity(data_frame: pd.DataFrame, p1_name: str, p2_name: str, at_frame: int):
+        p1_vel_mag = np.sqrt(data_frame[p1_name].vel_x.loc[at_frame] ** 2 +
+                             data_frame[p1_name].vel_y.loc[at_frame] ** 2 +
+                             data_frame[p1_name].vel_z.loc[at_frame] ** 2)
+        p2_vel_mag = np.sqrt(data_frame[p2_name].vel_x.loc[at_frame] ** 2 +
+                             data_frame[p2_name].vel_y.loc[at_frame] ** 2 +
+                             data_frame[p2_name].vel_z.loc[at_frame] ** 2)
+        # Check if initial player velocities are below MIN_BUMP_VELOCITY.
+        if all(x < MIN_BUMP_VELOCITY for x in [p1_vel_mag, p2_vel_mag]):
+            return False
+        else:
+            return True

carball/analysis/events/event_creator.py

@@ -92,7 +92,7 @@ def create_bumps(self, game: Game, proto_game: game_pb2.Game, player_map: Dict[s
                      data_frame: pd.DataFrame):
         logger.info("Looking for bumps.")
         bumpAnalysis = BumpAnalysis(game=game, proto_game=proto_game)
-        bumpAnalysis.get_bumps_from_game(data_frame)
+        bumpAnalysis.get_bumps_from_game(data_frame, player_map)
         logger.info("Found %s bumps.", len(proto_game.game_stats.bumps))
 
     def create_dropshot_events(self, game: Game, proto_game: game_pb2.Game, player_map: Dict[str, Player]):

carball/tests/stats/bump_test.py

@@ -0,0 +1,16 @@
+from carball.tests.utils import get_raw_replays, run_analysis_test_on_replay
+
+from carball.analysis.analysis_manager import AnalysisManager
+
+
+class Test_Bumps:
+    def test_calculate_bumps_correctly(self, replay_cache):
+        def test(analysis: AnalysisManager):
+            proto_game = analysis.get_protobuf_data()
+            count_bumps = 0
+            for i in proto_game.game_stats.bumps:
+                if not i.is_demo:
+                    count_bumps += 1
+            assert count_bumps == 7
+
+        run_analysis_test_on_replay(test, get_raw_replays()["7_BUMPS"], cache=replay_cache)

carball/tests/stats/demo_test.py

@@ -7,7 +7,10 @@ class Test_Demos:
     def test_calculate_demos_correctly(self, replay_cache):
         def test(analysis: AnalysisManager):
             proto_game = analysis.get_protobuf_data()
-            bumps = proto_game.game_stats.bumps
-            assert len(bumps) == 1
+            count_demo_bumps = 0
+            for i in proto_game.game_stats.bumps:
+                if i.is_demo:
+                    count_demo_bumps += 1
+            assert count_demo_bumps == 1
 
         run_analysis_test_on_replay(test, get_raw_replays()["1_DEMO"], cache=replay_cache)

carball/tests/utils.py

@@ -99,12 +99,17 @@ def get_raw_replays():
         "0_JUMPS": ["0_JUMPS.replay"],
         "0_SAVES": ["0_SAVES.replay"],
         "1_AERIAL": ["1_AERIAL.replay"],
-        "1_DEMO": ["1_DEMO.replay", "1_DEMO_1_63.replay"],
         "1_DOUBLE_JUMP": ["1_DOUBLE_JUMP.replay"],
         "1_EPIC_SAVE": ["1_EPIC_SAVE.replay"],
         "1_JUMP": ["1_JUMP.replay"],
         "1_NORMAL_SAVE_FROM_SHOT_TOWARD_POST": ["1_NORMAL_SAVE.replay"],
 
+        # Bumps and Demos
+        "1_DEMO": ["1_DEMO.replay", "1_DEMO_1_63.replay"],
+        "3_BUMPS": ["3_BUMPS.replay"],
+        "4_BUMPS": ["4_BUMPS.replay"],
+        "7_BUMPS": ["7_BUMPS.replay"],
+
         # Boost
         "3_STEAL_ORANGE_0_STEAL_BLUE": ["3_STEALS.replay"],
         "12_BOOST_PAD_0_USED": ["12_BOOST_PAD_0_USED.replay"],